Skip to main content

Molecular Manufacturing Systems

14.1. Overview

To bridge the gap between processing molecules and delivering macroscopic products, several issues must be considered. These include:

  • Large-scale assembly operations
  • Delivery to an external environment
  • Cooling requirements
  • Information sources
  • Control mechanisms
  • System architectures

Section 14.2 describes assembly operations at intermediate scales. Section 14.3 examines issues of system architecture, including overall organization, delivery of products, and reliability and redundancy. Section 14.4 puts these together to describe a system capable of manufacturing kilogram-scale product objects with a wide range of structures, including further manufacturing systems. Section 14.5 compares and contrasts present manufacturing technologies to molecular manufacturing technologies, considering their feedstocks, products, byproducts, energy consumption, productivity, quality, and cost. Finally, Section 14.6 examines the complexity of macroscopic products of molecular manufacturing, including the manufacturing systems themselves, drawing on experience with complexity in computer design and software.

14.2. Assembly operations at intermediate scales

Several considerations (assembly speed, the mass of products in transit, the size of assembly units in fault-tolerant architectures) favor the use of convergent assembly over moiety-by-moiety assembly in making product structures of substantial size and complexity. In convergent assembly, small parts (initially individual transferable moieties) are combined to make larger parts, and these to make still larger parts, in a hierarchically organized process. Convergent assembly is familiar in macroscale manufacturing processes as well: in making an automobile, parts are combined to make engines before engines are joined to the chassis; electronic components are combined to make circuit boards before boards are installed into radios, and radios are assembled before being joined to the dashboard.

Section 14.4 will describe a flexible manufacturing architecture based on convergent assembly. Given moiety-level assembly, the chief physical questions raised by this approach involve assembly at intermediate scales.

14.2.1. Joining building blocks

a. Mechanisms for positioning and assembly. Both mill-style and manipulator-style mechanisms can be used to transport and place building blocks of intermediate size. Section 13.4.4 discusses manipulators of sizes intermediate between the 100 nm100 \mathrm{~nm} and macroscopic scales; systems of belts, rollers, and cams can also be built at intermediate scales, again providing increases in mechanical stiffness.

The scale of these mechanisms depends on the scale of the parts being handled and of the product being assembled. In mill-style mechanisms, the transport system components must be scaled to the input parts. In manipulator-style mechanisms, the range of motion and hence the size of the device must be scaled to the output product.

Some useful assembly operations at intermediate scales may be quite specialized, such as stretching the sleeve of a grooved sleeve bearing in order to fit it over a shaft. Although specialized mechanisms are not explicitly considered here, they are likely to be commonplace in practice.

b. Interfaces between blocks. The assembly of blocks to form larger structures spans a continuum of sizes between cluster-based synthesis strategies (Section 8.6.6) and joining of macroscopic blocks. At sufficiently large scales (e.g., 100 nm\sim 100 \mathrm{~nm} ) the use of snaps, screws, and the like is feasible. On all scales above the smallest, adhesive interfaces of the sort described in Section 9.7 are applicable. As noted, some adhesive interfaces can join to form a dense, continuous diamondoid structure (or diamond itself). As a consequence, designing systems to be built as compositions of many small parts need not entail any great sacrifice in structural strength and stiffness.

Systems of molecular machinery will commonly consist of many distinct, free-standing parts, joined by interlocking geometries rather than by covalent bonds. Systems of this sort can be built by methods resembling conventional robotic assembly, using grippers that do not bond to the part being moved. Surface forces (e.g., van der Waals attraction) must be considered in designing grippers and assembly sequences, but these present no great difficulty where diamondoid structures with nonreactive surfaces are concerned, given assembly adhesion. (The construction of free-standing parts can be accomplished by building up a workpiece on a substrate in a manner that allows clean cleavage at the workpiece-substrate interface; the cleaved surface can then be finished or extended by further mechanosynthetic operations. The forces required for cleavage and subsequent manipulation can be provided by nonbonding grippers.)

c. Energy efficiency. The energy efficiency of intermediate-scale assembly can be high, on a per-atom basis. In moving large blocks, internal friction in positioning mechanisms is amortized over many atoms. A typical upper bound on the energy dissipated in joining surfaces is comparable to the surface energy, which for a large, thick block is small on a per-atom basis.

Energy dissipation is accordingly dominated by assembly of the smallest blocks. These can be approximated as one-nanometer cubes (consistent with the system description in Section 14.4). If the energy dissipated in joining surfaces were equal to the theoretical surface energy of diamond [5 J/m2\left[\sim 5 \mathrm{~J} / \mathrm{m}^{2}\right. (Field, 1979)], this would amount to 9×106 J/kg\sim 9 \times 10^{6} \mathrm{~J} / \mathrm{kg}. It is reasonable to expect, however, that much or almost all of the potential energy of bonding can be recovered as mechanical work, reducing dissipation to a small fraction of this value. If adhesive interfaces were designed such that the exoergicity of CC\mathrm{C}-\mathrm{C} bond formation matches that in the polymerization of ethene, then the exoergicity of assembling a diamondoid solid from cubic nanometer blocks would be 1×106 J/kg\sim 1 \times 10^{6} \mathrm{~J} / \mathrm{kg}. This value probably overestimates the total energy dissipated in block assembly in a well-designed system producing a mix of diamondoid products; the estimate used here is 5×105 J/kg5 \times 10^{5} \mathrm{~J} / \mathrm{kg}.

14.2.2. Reliability issues

In the assembly of intermediate-scale building blocks, ample exoergicity for stable, reliable assembly is no problem; even van der Waals attractions yield >150maJ>150 \mathrm{maJ} for typical interfaces of 1 nm2\geq 1 \mathrm{~nm}^{2}. Further, the greater stiffness of larger structures makes it increasingly easy to limit positional uncertainty resulting from thermal excitation to low, acceptable values. The special problems arising on a larger scale are greater sensitivity to nonthermal vibration resulting from an increased ratio of mass to stiffness, and greater radiation damage rates resulting from increased mass directly.

a. Vibration and alignment. Under a fixed acceleration, inertial loads cause elastic deformations that scale in proportion to the square of the characteristic length of a system. Oscillating accelerations (vibrations) produce dynamic deformations that depend on resonant frequencies and damping. For perspective, a manipulator 10 cm10 \mathrm{~cm} long and scaled from the device described in Section 13.4 has a length and stiffness greater by a factor of 10610^{6} and a mass greater by a factor of 101810^{18}. It deforms by 10 nm\sim 10 \mathrm{~nm} under terrestrial gravity, and its lowest resonant frequency is >104 Hz>10^{4} \mathrm{~Hz}.

Various combinations of (1) assembly tolerances, (2) measurement and control, and (3) precomputed corrections can compensate for static deformation. Various combinations of (1), (2), and mechanical isolation can correct for vibration, under ordinary circumstances. A surprise in the development of scanning tunneling microscope technology was the feasibility of achieving atomic-scale positional stability in centimeter-scale mechanisms mounted on ordinary laboratory tables, using 10cm10-\mathrm{cm}-scale vibration isolation systems consisting of ordinary elastomers and metal plates. Engineering molecular manufacturing systems for isolation from external vibrations and for control of internal vibrations should be feasible.

At larger scales, assembly tolerances can be increased without sacrificing precise control of the structure and alignment of the resulting interfaces. Figure 14.1 illustrates an adhesive interface structure in which nonbonded contacts occur as tapered pegs are inserted into matching holes. Tapering can provide substantial tolerance for initial misalignment, yet as the adhesive surfaces themselves approach contact, a stiff alignment can be achieved. As shown in the

Figure 14.1. An adhesive interface with tapered alignment pegs and a geometry that prevents residual nonbonded interfaces from serving as crack initiation sites.

illustration, interfacial bonding can proceed outward from the alignment pegs, and the structure around the edge of the nonbonded contact surface can be placed in compression to inhibit crack initiation.

b. Radiation damage: beyond the single-point failure assumption. Typical nanometer-scale machine components have many locations where a point defect induced by ionizing radiation can cause failure; the default assumption throughout this work has been that a defect at any point will cause failure. Macroscale machines, however, function despite numerous defects. At some intermediate scale, machines can (and must) be designed such that a significant density of point defects is tolerable. This scale can be estimated by considering the effects of radiation damage on two basic characteristics of mechanical parts: their shapes and the properties of their interfaces. (Strength, stiffness, and the like are not at issue in the typical case.)

Applying an elastic continuum model, the effect of a point defect on the shape of a component can be described in terms of the resulting strain field. This is in general complex, but its longest-range component results from the local change in volume. Modeling the medium as homogeneous, isotropic, and elastic, the resulting radial strain Δr\Delta r falls off as r2r^{-2}. An atomic-scale defect causes a strain of Δr<0.2 nm\Delta r<0.2 \mathrm{~nm} at r=0.2 nmr=0.2 \mathrm{~nm}, and hence a strain Δr<0.002 nm\Delta r<0.002 \mathrm{~nm} at 2 nm2 \mathrm{~nm}, and so forth. Since geometrical errors of 0.01 nm\sim 0.01 \mathrm{~nm} are tolerable in most mechanosynthetic systems, components having features of multinanometer or larger scale should usually prove able to tolerate radiation-induced defects in their interiors without failure. Large structural components with damage-sensitive interfaces (such as the segments of the stiff arm described in Section 13.4) can have substantial damage-tolerant interior and surface regions, in effect reducing their target sizes relative to the model described in Section 6.6.2.

Similar principles apply to misalignments caused by damage to bearing interfaces, although the nonlinearity of nonbonded interactions tends to increase the characteristic length scales. Moreover, damage occurring at a mobile interface can in some instances produce large static friction or adhesion, causing device failure. Bearings can be made damage tolerant by providing internal redundancy. For example, the surfaces of a sleeve bearing can be separated by multiple independent layers, each subdivided into parallel strips. A structure of this sort can tolerate several adhesion points among adjacent surfaces, strips, and layers without creating a connection between the sliding surfaces. Designs of this sort typically require interfacial thicknesses of several nanometers, which are geometrically compatible only with moving parts on a scale of tens of nanometers or more. Thus, systems large enough to employ internally redundant bearing interface structures commonly have structural components large enough to maintain good geometrical stability in the presence of point defects.

In a mill-style manufacturing process without fault-correction, a failed assembly operation can cause failure of a relatively large subsystem. The estimates made in this section regarding the scale at which machine components can tolerate point defects without failure suggest that assembly processes can tolerate damage in the parts themselves when these are of a similar scale. In the presence of point defects, assembly can proceed without disrupting the process; whether the resulting product is functional is a separate question. Accordingly, it will be assumed that when the scale of all components and building blocks in a manufacturing process is >10 nm>10 \mathrm{~nm}, a moderate density of point defects can be tolerated without causing failure of the process.

14.3. Architectural issues

The previous sections of this chapter have chiefly described devices, subsystems, and operations, covering molecular sorting and orientation, transformation of molecules into reagent moieties, the application of reagent moieties to build small structures, and the hierarchical assembly of smaller structures into progressively larger structures. The present section outlines an architecture for molecular manufacturing systems capable of producing a wide range of macroscopic products and delivering them to an external environment. This requires attention to the spatial organization of material and energy flows, and to such system-level concerns as overall reliability and operational lifetime. These are addressed in substantial detail in the description of an exemplar architecture.

14.3.1. Combining parts to make large systems

Present manufacturing practice provides no precedent for processes that combine 1025\sim 10^{25} distinct parts to form a single object. A naive approach would use a mechanism to assemble parts one at a time; at 10610^{6} per second, however, assembling 102510^{25} parts would take a time longer than standard estimates of the age of the universe. Viable approaches must coordinate many mechanisms in building a single object.

a. Construction-style assembly processes. A construction-style process uses small devices to work on or in a large structure. For example, many assembly mechanisms can work in parallel to build up a surface. If a product structure is 0.1 m0.1 \mathrm{~m} thick, and each assembly mechanism occupies an area of (100 nm)2(100 \mathrm{~nm})^{2}, then an assembly rate of 10610^{6} small moieties per second implies a 108 s\sim 10^{8} \mathrm{~s} construction time, more than a year. This strategy might be attractive for designs that lack thick solid regions: various porous, fibrous, foamlike, or honeycombed structures meet this criterion and have attractive mechanical properties.

Alternatively, assembly can proceed at higher rates through the use of larger building blocks. Keeping other parameters constant, increasing the block size to 10 nm\sim 10 \mathrm{~nm} decreases the assembly time to 1000 s1000 \mathrm{~s}. Blocks of sufficient size and complexity (e.g., containing a computer, motor, and actuators) could be made selfassembling, although at a substantial penalty in properties such as strength-todensity ratio.

Some products can doubtless be assembled in a poorly controlled environment (e.g., in a solvent bath), simplifying problems of heat and mass transport, and of environmental control. Alternatively, eutactic environments of almost any desired size can be constructed by expanding a gas-tight barrier in the manner shown in Figure 14.2 or Figure 14.3. This style of assembly will not be considered further here.

b. Manufacturing-style assembly processes. In manufacturing-style processes, parts are manipulated and transported within larger mechanisms. If a construction-style process resembles the assembly of a building, a manufacturing-style process resembles the assembly of a computer. Processes of this sort will be described in greater depth.

The architecture described in Section 14.4 uses a convergent assembly sequence in which each structure is built from components within an order of magnitude or so of the structure's own linear dimensions. In one class of convergent assembly processes, the motion of components traces a tree in space: the trunk corresponds to the path traced by the final workpiece as the final components are assembled, the branches correspond to the paths traced by those components as they are assembled, and so forth. A simplified model (Figure 14.4) demonstrates that convergent assembly can be distributed in space in a manner that (1) provides an assembly volume proportional in size to the workpiece at each stage, and (2) requires only short-range transportation of parts between stages.

A system of the sort suggested by Figure 14.4 must operate all assembly stations at the same frequency, thereby slowing the smallest mechanisms to the frequencies of the largest. Varying the number of assembly operations per level in the hierarchy can alleviate this problem, as can allowing a small number of highfrequency processes to supply parts to a larger number of low-frequency processes. Both of these strategies are incorporated into the exemplar architecture summarized in Section 14.4.

The boundary between high-frequency and low-frequency processes in the exemplar architecture corresponds to a boundary between mill-style and manipulator-style assembly mechanisms. This, in turn, marks the boundary between purely repetitive operations producing standard building blocks and programmable operations that can stack these building blocks to make a wide variety of products. Although manipulators can handle blocks of 10 nm\sim 10 \mathrm{~nm} scale with good energy efficiency, the exemplar architecture uses mill-style mechanisms up to a 1μ1 \mu scale. This sacrifices flexibility, but the number of distinct kinds

Figure 14.2. Expansion of a eutactic environment by construction mechanisms operating inside it (ports for transfer of materials across the boundary not shown). In three dimensions, the arrangement of folds at the corners where three edges meet is more complex than suggested in steps 2 or 3 .

of 1μ1 \mu blocks can be quite large. A 1 kg1 \mathrm{~kg} structure contains 1015\sim 10^{15} blocks. To manufacture them, the exemplar architecture uses 106\sim 10^{6} separate systems, each capable of producing a different structure. The demand for large numbers of a few kinds tends to reduce diversity, but the option of running some lines at low capacity (or only for producing certain products) substantially offsets this. The

Figure 14.3. Expansion of a eutactic environment as in Figure 14.2, using sliding blocks rather than flexible pleats. This approach is based on the presumption that adequate seals can be maintained at all the junctions among the sliding blocks. Again, geometries become more complex with expansion in three dimensions, but local blocksliding relationships are unaltered. Note that the "blocks" could in principle be atoms of a metallic material, with sliding occurring by dislocation movement.

Figure 14.4. A simple model of a spatial arrangement for a hierarchical, convergent assembly process; panels (a), (b), (c), and (d) provide successively more detailed diagrams. In (a), a primary assembly line consisting of a series of 8 assembly stations (drawn as cubes) performs the final 8 assembly operations in a hypothetical manufacturing process. In (b), 8 secondary assembly lines provide parts to the final lines; (c) and (d) illustrate tertiary and quaternary assembly lines. Since each level of lines contains an equal volume, this pattern cannot be indefinitely extended without self-intersection (the maximum radius of expansion is bounded). With local rearrangements to postpone self-intersection until the available volume is nearly full, a branching pattern of this sort can be extended though >30>30 generations, enabling the assembly of objects from >1027>10^{27} pieces. This structure demonstrates that certain geometrical constraints can be met, but does not represent a proposed system.

range of systems that can be constructed from a catalogue of 106\sim 10^{6} kinds of parts is large, particularly when the parts can be joined to make continuous volumes of high strength materials, and are each smaller than the usual tolerances in conventional manufacturing.

14.3.2. Delivering products to an external environment

In typical applications, the products of molecular manufacturing must be delivered to an external environment without permitting back contamination of the eutactic internal environment. Sections 11.4.2 and 11.4.3 discuss vacuum pumps and the use of sliding interfaces as seals; from these elements, various means can be devised for product delivery.

For example, a product object can be placed in a box bearing multiple external sealing rings and followed by a similarly ringed piston. The rings can provide a tight seal against the inside of an exit tube, and the piston can block the exit until another product-bearing box is ready for departure. Moderate failures in the seals could be rendered harmless by scavenging pumps operating on the spaces between them. Schemes that lack discardable boxes, pistons, and the like


Figure 14.5. Schematic illustration of a gas-tight, expandable, doubly pleated enclosure for a eutactic environment (a). In three dimensions, a structure like that illustrated could undergo a volumetric expansion by a factor of 27\sim 27, unrolling one band of pleats at a time. Note that each face must have pleats running in two perpendicular directions, as shown in (b), an angled view of the surface of a doubly pleated enclosure. To satisfy constraints on elastic strain, the minimum radius of curvature in the first set of pleats must be several times the thickness of the gas-tight wall, and the radius of curvature in the second, superimposed set of pleats must be a further multiple of the radius of the first. Accordingly, the enclosure size must be orders of magnitude greater than the wall thickness, making this approach inapplicable to very small enclosures. Note that the wall system can include structural members and sturdy barriers both inside and outside of the comparatively delicate gas-tight wall.

can also be devised, with varying volume requirements, constraints on product object shape, and so forth. These will not be pursued further at present. Figures 14.5 and 14.6 illustrate how delivered products can be larger than the system that produced them.

14.3.3. Redundancy, reliability, and system lifetimes

A system of macroscopic scale and significant lifetime must tolerate localized, randomly distributed damage, which in nanomechanical systems will commonly cause failure of devices and subsystems. There is a large and diverse literature on fault-tolerant design for computers and aerospace systems, and similar approaches are applicable in many nanomechanical systems.

With a judicious choice of structures and reactions, damage to a molecular manufacturing system at ordinary temperatures is dominated by damage caused

Figure 14.6. Schematic illustration of a geometry (suggested by R. Merkle) permitting the manufacture and delivery of objects of a size equaling or exceeding (in all dimensions) that of the manufacturing system.

by ionizing radiation. Accordingly, Eq. (6.54) can be taken as a damage model. Using this model, Section 13.3.6c concludes that mechanosynthetic units with a mass of 2×1018 kg\leq 2 \times 10^{-18} \mathrm{~kg} have a probability of failure .01\leq .01 in a 10 -year period in a typical terrestrial environment. The early sets of mill units described in Section 14.4 have overall operating frequencies of 106 s1\sim 10^{6} \mathrm{~s}^{-1}, but divided among 100 units in each set, yielding an operating frequency of 104 s110^{4} \mathrm{~s}^{-1} per unit. Accordingly, if the failure rate of a mill unit is dominated by a step with a failure rate of 101510^{-15} (see Section 13.3.6), this will contribute .003\sim .003 to the probability of failure in a 10 -year period.1

Section 13.3.6b shows that mechanosynthetic units can be built to exhibit fail-stop behavior, reliably ceasing operation rather than delivering damaged products. Section 13.3.2b outlines the characteristics of transportation components capable of transferring structures from belts to pallets, moving pallets along tracks through fair, nonblocking merging and distribution junctions, and transferring structures back to belts again. Figure 14.7 shows how fail-stop units

Figure 14.7. Schematic diagram of a portion of a redundant, fault-tolerant manufacturing architecture. The upper set of identical, fail-stop producers transforms inputs into intermediate products that become inputs to a set of identical, fail-stop consumers. The intermediate transportation network incorporates fair, nonblocking merging junctions (m) and distribution junctions (d). A system with this structure can continue to operate despite the failure of any component, including the failure of all but one producer and all but one consumer. Within the pattern illustrated, the degree of redundancy of producers, consumers, and transportation elements can be increased arbitrarily.

can be joined by redundant tracks and junctions in such a manner that any single unit, track segment, or junction can fail without interrupting the transformation of inputs into outputs. In particular, the system can continue to work without error so long as one unit in any set remains operational. (The presence of redundant units decreases the operational frequency of each, decreasing energy dissipation per operation.)

In the typical case, failure rates are dominated by failures in the mechanosynthetic units rather than in the transportation network. In this case, the probability Psyst P_{\text {syst }} that a system remains operational is related to the probability Punit P_{\text {unit }} that a unit remains operational by the expression

Psyst =[1(1Punit )Nunit ]Nset \begin{equation*} P_{\text {syst }}=\left[1-\left(1-P_{\text {unit }}\right)^{N_{\text {unit }}}\right]^{N_{\text {set }}} \tag{14.1} \end{equation*}

where Nunit N_{\text {unit }} and Nset N_{\text {set }} are the number of units in a set and sets in a system, respectively. (Unit failures are assumed to be independent; substantial spatial separations can be provided.) A more readily evaluated expression is

Psyst exp[Nset (1Punit )Nunit ], where Punit 1\begin{equation*} P_{\text {syst }} \approx \exp \left[-N_{\text {set }}\left(1-P_{\text {unit }}\right)^{N_{\text {unit }}}\right], \quad \text { where } \quad P_{\text {unit }} \approx 1 \tag{14.2} \end{equation*}

These relationships correspond to Eqs. (6.58) and (6.59).

14.4. An exemplar manufacturing-system architecture

14.4.1. General approach

This section and Table 14.1 outline an architecture for a system capable of manufacturing macroscopic objects. The subsystem capacities are chosen to permit the conversion of a feedstock solution consisting of small organic molecules into 1 kg\sim 1 \mathrm{~kg} product objects of 0.2 m\sim 0.2 \mathrm{~m} dimensions in a cycle time of 1\sim 1 hour. The flow of materials proceeds through molecule sorting and orientation (Section 13.2), preparation of reagent moieties (Section 13.3), several stages of convergent assembly using mill-style mechanisms (Sections 13.3 and 14.2), and several stages of convergent assembly using manipulator-style mechanisms (Sections 13.4 and 14.2).

This section focuses on mechanosynthesis and assembly, omitting the details of supporting systems. For example, although they would be essential in a detailed design, the specifics of a 1 kW\sim 1 \mathrm{~kW} cooling system, a0.001 kg/sa \sim 0.001 \mathrm{~kg} / \mathrm{s} feedstock solution supply system, transportation and sealing mechanisms for product delivery, and so forth are peripheral to the central issues of molecular manufacturing. Figure 14.4 shows that convergent assembly processes can be performed without undue geometrical problems in material flow; accordingly, a reasonable estimate of overall system volume can be had by summing the volumes of the assembly workspaces without describing a particular three-dimensional layout. Maintenance of vacuum integrity has been discussed in Sections 11.4.2, 11.4.3, and 14.3.2. Compressive structures made of diamondlike materials can support terrestrial atmospheric pressures with masses of 1 kg/m3\ll 1 \mathrm{~kg} / \mathrm{m}^{3}, and can be designed in many ways. Power conversion between mechanical and electrical forms can be performed with high efficiency and (for reasonable power levels) with negligible mass and volume (Section 11.7), imposing few constraints in this context.

The key issues in a manufacturing architecture of this sort center around the assembly mechanisms themselves: their numbers, kinds, reliabilities, operating frequencies, masses, volumes, and so forth. These issues are addressed in the next section; many parameters are summarized and explained in Table 14.1.

14.4.2. Products, building blocks, and assembly sequences

Using fixed, mill-style subsystems throughout a manufacturing system, from the moietal level to finished products, would limit it to making products of a single kind. Using flexible, manipulator-style subsystems throughout would permit independent control of every detail of a product, permitting an indefinitely large range of structures at all scales, but at the cost of relatively slow, inefficient manufacturing processes. Neither of these extremes is likely to prove desirable.

a. Building blocks and product diversity. Typical macroscopic products can conveniently be made from modular building blocks, many identical to one another. Micron-scale building blocks are small enough to make almost any macroscopic shape in ordinary use today within better tolerances than those provided by conventional machining. From sets of building blocks that can be assembled to make solids of differing strength, stiffness, conductivity, and so forth, objects can be made that exhibit material properties that equal or better those of present industrial products. Systems that take more direct advantage of nanoscale structures can still use extensively duplicated building blocks. Products containing a macroscopic quantity of computational hardware will typically contain many identical CPUs, memory arrays, and so forth. A product containing a macroscopic quantity of mechanical structure will typically contain many identical regions of structural material. Motors, gearboxes, vacuum pumps, databus segments, photovoltaic cells, storage batteries, conductors, struts, connectors, manipulator arms, reagent-processing subsystems-all can be constructed on a micron or submicron scale, and all are candidates for extensive duplication in macroscopic systems.

Accordingly, a wide range of macroscopic products can be made with good efficiency by using mills to make a diverse set of building blocks in the 10710^{-7} to 106 m10^{-6} \mathrm{~m} size range and then using manipulators to assemble them into macroscopic products. Inclusion of a secondary production capability that applies manipulators to smaller building blocks (at the moietal level where necessary), can provide unusual building blocks without requiring a dedicated mill mechanism. The exemplar system in Table 14.1 makes the simplifying assumption of a hand-off from mills to manipulators at a single block size, 106 m10^{-6} \mathrm{~m}.

b. Alternative assembly sequences. The quantitative summary of the exemplar architecture assumes a strictly convergent assembly sequence in the mill mechanism: each moiety and each subassembly is destined for a unique product at the 106 m10^{-6} \mathrm{~m} block level. The reagent preparation mill system, in contrast, produces standard moieties that can be used by any of a wide range of assembly mills. There is good reason to extend this pattern to the subassembly level, thereby permitting flexible allocation of the output of subassembly mill mechanisms to any of several higher-level assembly mills. Doing so enables a system to produce a greater range of 106 m10^{-6} \mathrm{~m} scale blocks with a lesser increase in total mill system size.

A further assumption in the quantitative summary is that each assembly unit combines blocks of similar sizes to make a much larger structure. There is no reason to treat this as a constraint, and there are substantial advantages to assembling components of widely differing sizes. For example, after assembling two large blocks with an adhesive interface, moietal-scale modification of the edge of the interface region might be necessary to prepare a smooth surface spanning the two blocks. Macroscopic assembly processes frequently add small components to large structures.

c. Adjustable blocks. A large number of blocks will serve simple structural functions, holding active components such as motors, gearboxes, sensors, and the like in fixed geometrical relationships to one another. This requires struts of diverse lengths and joints of diverse angles. There is, however, no reason why structural elements of each needed length and angle must be manufactured by a distinct assembly mill, starting at the moietal level. It is straightforward to build struts and joints that contain internal sliding interfaces (Section 10.5) permitting them to be extended or twisted to assume a wide range of lengths or angles, and then twisted or compressed (respectively) to bring internal adhesive interfaces (Section 9.7) into contact, thus producing a stiff structural member with a particular geometry. In this manner, a mill system can make identical, adjustable members that each of several subsequent mills (or manipulators) can transform into one of many fixed, specialized structures.

14.4.3. Throughput, delays, and internal inventories

If all components operated at their stated frequencies, handling components with a density of 1000 kg/m31000 \mathrm{~kg} / \mathrm{m}^{3}, system throughput would be 0.001 kg/s0.001 \mathrm{~kg} / \mathrm{s}. The five mill stages operate at high frequencies, performing relatively small numbers of operations per output object; the maximum delay for material passing through these five stages is 0.1 s\sim 0.1 \mathrm{~s}, implying a 104 kg\sim 10^{-4} \mathrm{~kg} inventory of materials in process. The three manipulator stages operate at lower frequencies, performing relatively large numbers of operations per output object; each requires 1000 s1000 \mathrm{~s} to complete its operations, and contains a time-average product mass of 0.5 kg0.5 \mathrm{~kg}. Allowing for feasible overlap in the operating times of the manipulator stages, the delay between process initiation and delivery of a 1 kg1 \mathrm{~kg} product object, starting from a feedstock solution, can be somewhat less than one hour.

14.4.4. Mass and volume

The volume of the first stages-input ordering, reagent preparation, and mill mechanisms - can be estimated by applying an estimated mean density to the tabulated masses. Using 100 kg/m3100 \mathrm{~kg} / \mathrm{m}^{3} yields a volume of 1.8×104 m3\sim 1.8 \times 10^{-4} \mathrm{~m}^{3}. (System flexibility could be increased by devoting 10\sim 10 times this mass and volume to mill mechanisms, most idle in a typical production process; other system parameters would remain essentially unchanged.)

The volume of the manipulator stages depends on the size of the product structures, if these are to be totally enclosed in the workspace rather than constructed and extruded incrementally. A wide range of 1 kg1 \mathrm{~kg} products should fit (perhaps in a folded or partially disassembled configuration) within a 0.01 m30.01 \mathrm{~m}^{3} workspace. If the solid parts of a 1 kg1 \mathrm{~kg} product object are as dense as diamond, then these parts occupy 0.03\sim 0.03 of the workspace volume. Assuming that workspace volumes are nonoverlapping, but can be used for conveyance of parts from other workspace volumes, then the total volume required for the three manipulator stages is 0.03 m3\sim 0.03 \mathrm{~m}^{3}. Multiplying by 1.66 to allow for packing inefficiencies, cooling channels, and so forth yields a total estimated volume of 0.05 m3\sim 0.05 \mathrm{~m}^{3}.

The total mass of the manufacturing system components listed in Table 14.1 is <0.12 kg<0.12 \mathrm{~kg}. Allowing 1 kg/m31 \mathrm{~kg} / \mathrm{m}^{3} for internal compressive structure to support external atmospheric pressure adds 0.05 kg\sim 0.05 \mathrm{~kg}; allowing 0.1 kg/m20.1 \mathrm{~kg} / \mathrm{m}^{2} to provide for a sturdy case, rubber feet, and the like adds 0.08 kg\sim 0.08 \mathrm{~kg}. These masses sum to <0.25 kg<0.25 \mathrm{~kg}. Allowing as much as 0.75 kg0.75 \mathrm{~kg} for other subsystems yields an estimated total system mass <1 kg<1 \mathrm{~kg}.

14.4.5. System lifetime

Radiation damage chiefly threatens the small yet relatively specialized devices in the first three stages of the mill system. Earlier stages are massively redundant; later stages use components large enough to permit radiation-insensitive designs (Section 14.2.2b). Applying the previous radiation-damage assumptions, units with a radiation-sensitive mass of 2×1018 kg\leq 2 \times 10^{-18} \mathrm{~kg} have Punit =.99P_{\text {unit }}=.99 after ten years. The radiation-sensitive portions of mill stages 1-3 have a lower unit mass, with Nunit =10,Nset 1015N_{\text {unit }}=10, N_{\text {set }} \approx 10^{15}, and hence Psyst 1P_{\text {syst }} \approx 1 after ten years. Increasing Nunit N_{\text {unit }} to 20 raises the expected lifetime to over a century. All radiation-sensitive units are assumed to be fail-stop (Section 13.3.6) and linked by redundant transport networks (Section 13.3.2b). An enclosure of this size can easily be made UV opaque (Section 6.5.4).

14.4.6. Feedstock materials

A specific manufacturing mechanism will accept raw materials consisting of a specific set of small molecules in solution. Typical products require large quantities of C\mathrm{C}, moderate quantities of H,N,O,F,Si,P,S\mathrm{H}, \mathrm{N}, \mathrm{O}, \mathrm{F}, \mathrm{Si}, \mathrm{P}, \mathrm{S}, and Cl\mathrm{Cl}, and lesser quantities of several other elements. These could be provided by a single solution, or by different solutions supplied through several ports. In designing a system, the choice of input compounds is largely a matter of cost and convenience.

14.4.7. By-products

Extended diamondlike structures have a smaller mass fraction of hydrogen than do typical small organic molecules: small molecules have a larger surface-to-volume ratio, and satisfaction of surface valences is most commonly achieved by inclusion of monovalent atoms, chiefly hydrogen. Fullerenes, cyanogen, perfluoroalkanes, and other unusual molecules are among the exceptions to this pattern. Aside from this, there is no strong reason why the elemental composition of the inputs should differ from that of the product, and hence no strong reason why there should be substantial by-products from a molecular manufacturing system. Excess hydrogen could be delivered as water, in a process that consumes atmospheric oxygen and (presumably) produces mechanical energy as a by-product (Section 13.3.8).

If a mechanism is supplied with a single solution of fixed composition, and must from this produce a variable mix of products, the (nonhydrogen) elemental composition of the products will in general fail to match that of the input solution, and the less-consumed compounds will accumulate. The resulting solutions can be recycled by adding the compounds in which they have become deficient.

14.4.8. Energy output and dissipation

The energy dissipated by mills is dominated by moietal operations, estimated by Section 13.3 .7 as 1.5×106 J/kg1.5 \times 10^{6} \mathrm{~J} / \mathrm{kg}; a further allowance of 5×105 J/kg5 \times 10^{5} \mathrm{~J} / \mathrm{kg} is made for energy dissipated during block assembly (Section 14.2.1c). The mechanical aspects of manipulator operations result in comparatively negligible energy dissipation, as do the operations in molecular sorting and ordering (save for entropic effects included in the overall thermodynamic calculations that follow).

The energy dissipation resulting from computation used to direct manipulator operations is hard to estimate accurately, but can be made quite small in the exemplar architecture. The control of motions can be relatively direct: with compiled instructions, no run-time planning is necessary, hence the computations can be as simple as loading appropriate internal-coordinate settings into registers in controllers. A relatively computation-intensive process (perhaps analogous to the interpretation of a page-description language) might execute 106\sim 10^{6} instructions per block placed. Section 12.7.4 derives an estimated energy dissipation of 1016 J\sim 10^{-16} \mathrm{~J} per instruction, and a 1 kg1 \mathrm{~kg} object contains 1015\sim 10^{15} cubic-micron blocks; hence executing even 10610^{6} instructions per block would increase the total energy dissipated by only 105 J/kg10^{5} \mathrm{~J} / \mathrm{kg}.

Transformation of hydrogen-rich organic feedstock molecules and oxygen into typical diamondoid products and water is an exoergic, entropy-reducing process. For example, a process with acetone as the primary carbon source and diamond as the chief product liberates a total energy equaling 1.7×107 J/kg\sim 1.7 \times 10^{7} \mathrm{~J} / \mathrm{kg} of useful output, while decreasing local entropy by 5.7×103 J/kgK\sim 5.7 \times 10^{3} \mathrm{~J} / \mathrm{kg} \cdot \mathrm{K} (approximating the entropy of the product object as zero). The change in free energy accordingly is 1.5×107\sim 1.5 \times 10^{7}. Allowing for 3.1×106 J/kg\sim 3.1 \times 10^{6} \mathrm{~J} / \mathrm{kg} of free energy dissipated (see preceding paragraphs), the system produces 1.2×107 J/kg\sim 1.2 \times 10^{7} \mathrm{~J} / \mathrm{kg} of surplus energy in the form of mechanical work. This could be conveniently disposed of by delivery to an external power distribution system in the form of electrical energy.

Waste heat from the process results from both inefficiencies and computation (3.1×106 J/kg)\left(\sim 3.1 \times 10^{6} \mathrm{~J} / \mathrm{kg}\right) and from the reduced entropy of the products relative to the feedstocks (1.7×106 J/kg)\left(\sim 1.7 \times 10^{6} \mathrm{~J} / \mathrm{kg}\right). Thus, a system producing 1 kg/hr1 \mathrm{~kg} / \mathrm{hr} of diamondoid products dissipates an estimated 1.3 kW1.3 \mathrm{~kW} of waste heat. The required cooling capacity can be provided by fan-driven air flowing at a rate of 0.1 m3/s0.1 \mathrm{~m}^{3} / \mathrm{s} with ΔT=\Delta T= <15 K<15 \mathrm{~K} between the intake and exhaust ports. Cooling channels can be placed in close proximity to the chief heat sources, permitting good thermal contact.

14.4.9. Information requirements

A typical product of the exemplar architecture consists of 1015\sim 10^{15} building blocks of 106 m\sim 10^{-6} \mathrm{~m} linear dimensions. To specify a product requires specifying the type and position of each block. If there are <109<10^{9} different kinds of block, then specifying the type of a block requires <32<32 bits of information (substantially less, with efficient encoding). The total quantity of information required per block need seldom be much larger than this.

An analogy can be drawn to a pattern of type on a page: letter forms are complex, but standardized; with a little additional information (e.g., column widths, line spacings, superscript and subscript codes) a sequence of 8-bit codes specifying a sequence of letters determines the two-dimensional arrangement of those letters on a page. This may, however, depend on large auxiliary tables of information on letter widths, spacings for kerned pairs, and so forth. In a similar manner, a sequence of codes specifying blocks can be combined with auxiliary information describing the acceptable geometrical relationships between adjacent blocks to provide most of the information required to determine the threedimensional arrangement of those blocks in space. This approach exploits stereotyped relationships among adjacent blocks to generate precise positioning information through simple computations.

Specification of an object should thus require substantially fewer than 100 bits per block, or <1017<10^{17} bits per product object. Since molecular tape storage systems can hold 1028\sim 10^{28} bits /m3/ \mathrm{m}^{3} (Section 12.6.4), a volume of <105 m3<10^{-5} \mathrm{~m}^{3}, containing <0.01 kg<0.01 \mathrm{~kg}, can contain sufficient information to specify the structures of 10610^{6} product objects. (Note that the presence of repeated substructures above the block level, either within or between objects, can greatly reduce information requirements.)

14.4.10. Manufacture of manufacturing systems

Since most of the internal volume of the exemplar architecture is devoted to open workspaces for manipulators, it should be feasible to design a system that can be folded from linear dimensions of <0.4 m<0.4 \mathrm{~m} to linear dimensions of 0.2 m\sim 0.2 \mathrm{~m}, or (more to the point) that can be unfolded from the smaller dimensions to the larger; Figure 14.5 illustrates how a continuous, gas-tight wall can be expanded by a greater factor. This places the exemplar architecture within both the size and mass range of its own products.

With the use of programmable manipulators to build a diverse set of structures from a smaller set of building blocks, the output of a set of specialized mills can be used to build an identical set of mills, as well as many other structures. Accordingly, a suitably designed and programmed molecular manufacturing system, built along the lines of the exemplar architecture, can be used to build objects that, when unfolded, are substantially identical to itself.

The design of machines able to construct copies of themselves was described in von Neumann and Burks (1966), and many variations on this theme have been reviewed in Freitas and Gilbreath (1982) and the included references. It may seem somehow paradoxical that a machine can contain all the instructions needed to make a copy of itself, including those selfsame complex instructions, but this is easily resolved. In the simplest approach (as implicitly adopted here), the machine reads the instructions twice: first as commands to be obeyed, and then as data to be copied. Adding more data does not increase the complexity of the data-copying process, hence the set of instructions can be made as complex as is necessary to specify the rest of the system. By the same token, the instructions transmitted in a replication cycle can specify the construction of an indefinitely large number of other artifacts.

Figure 14.8. The inputs and outputs of the exemplar manufacturing system discussed in Section 14.4.

14.5. Comparison to conventional manufacturing

The characteristics of molecular manufacturing systems are best understood in comparison to the characteristics of conventional manufacturing systems. The following comparisons embrace required inputs of materials and energy, byproducts, sizes of internal components, frequencies of internal operations, overall productivity, characteristics of the resulting products, and costs of production. Differences in the technology base required for implementation are profound, and are implicitly addressed in Chapters 15 and 16, which describe strategies for developing molecular manufacturing capabilities.

The discussion in this section is not a basis for further analysis of molecular machinery, manufacturing, or computation; rather, it attempts to evaluate some of the consequences of the previous analysis. In characterizing such broad topics as the entire range of product characteristics and the entire range of potential by-products, the descriptions are based on informed estimates. Most other conclusions are direct consequences of the previous analysis, supplemented by commonplace knowledge of present technological capabilities. Figure 14.8 summarizes the inputs and outputs of the exemplar manufacturing system.

14.5.1. Feedstocks and energy requirements

Defining what one means by a "feedstock" for conventional manufacturing can be difficult. A factory making products for household use may consume inputs as simple as sheet metal and wire, or as complex as integrated circuits and prefabricated assemblies. A system of factories considered as a whole typically includes smelters, steel mills, and the like, starting with crude feedstocks (iron ore, coal).

In molecular manufacturing, a mixture of simple compounds can serve as a feedstock (Sections 13.2 and 13.3). Quantitatively, the largest requirement is for a source of carbon. Agricultural products (sugar, alcohol) or petrochemicals can serve this function at costs of 0.1\textdollar/kg\sim 0.1 \textdollar / \mathrm{kg}, comparable to the cost of raw steel. These inputs are converted into finished products without intermediate stages that require external handling or transportation.

The discussion in Section 14.4.8 indicates that a molecular manufacturing process can be driven by the chemical energy content of the feedstock materials, producing electrical energy as a by-product (if only to reduce the heat dissipation burden). This contrasts with conventional manufacturing processes, which consume energy not contained in their inputs. Further, both the energy and material input requirements per unit of product functional capacity are substantially lower than those for conventional manufacturing processes, owing to greater functional capacity per unit product mass.

14.5.2. By-products and recycling

From the discussion in Sections 13.2 and 13.3, it should be clear that molecular sorting and processing mechanisms can perform transformations like those now performed in chemical plants, but with greater efficiency and control. The preparation of feedstock solutions need not generate by-products in uncontrolled forms; indeed, systems can be designed to produce no by-products at all, aside from some set of compounds containing any unwanted elements present in a crude input material. A general strategy for accomplishing this would be to extract and process small, soluble molecules from a mixture, then use conventional, nonspecific techniques (heat, oxidation, acids) to break down any residue into small, soluble molecules. Design studies to date suggest that relatively exotic and toxic materials (lead, mercury, cadmium) need play no role in molecular manufacturing processes or products.

Many products can be built as aggregates of modules joined by reversible means. These can presumably be recycled with little energy cost by disassembly and reassembly operating at the scale of micron-scale blocks. Failing this, almost any product can be recycled into small molecules subjecting it to chemical attack in a closed vessel (e.g., by oxidation in a sealed incinerator, followed by other processes to break down any residual solids).2 These small molecules can then be sorted and reprocessed.

14.5.3. Internal component sizes and frequencies

In conventional manufacturing, human hands and other devices on a 0.01 to 1 m1 \mathrm{~m} scale perform operations at frequencies that typically range from 10 to 0.1 Hz0.1 \mathrm{~Hz}. In molecular manufacturing, the most numerous operations are performed by devices on a 10910^{-9} to 107 m10^{-7} \mathrm{~m} scale, operating at frequencies of 106 Hz\sim 10^{6} \mathrm{~Hz}. In making macroscopic objects by convergent assembly, a modest number of operations are performed at frequencies of 10 Hz\sim 10 \mathrm{~Hz}.

14.5.4. Productivity

The physical productivity of a system can be measured in various ways. One is the time required for a system to produce outputs with a total mass equaling its own (by this standard, a short piece of pipe can be enormously productive). Another basis, less precise but still significant, is the length of time required for a system to produce outputs with a total complexity equaling its own, starting with simple inputs. If this criterion is applied to the materials processing and manufacturing sector of an industrial economy, the result corresponds to the minimum doubling time for the total capital stock, and is presumably on the order of a year or more, 108 s\sim 10^{8} \mathrm{~s}. For the exemplar molecular manufacturing system, the equivalent time is 103 s\sim 10^{3} \mathrm{~s}. This disparity stems from the difference in the characteristic frequency of operations.

14.5.5. Some feasible product characteristics

a. Precision and reliability. Most products made by present manufacturing systems are poorly ordered: they have either amorphous microstructures or crystalline microstructures that contain a high density of atomic defects, impurities, dislocations, grain boundaries, and microcracks. Semiconductor substrates come closest to perfection, but even so, most devices are formed by imprecise processes such as diffusion of impurities, nonepitaxial deposition processes, etching, and the like. In addition to having poorly controlled microstructures, macroscale objects have shapes that are rarely precise to <106 m<10^{-6} \mathrm{~m}. Without robust designs and careful process control, statistical variations in defect size can make products unreliable. Designs that tolerate material defects usually must sacrifice performance to do so.

Molecular manufacturing systems can make well-ordered products, save for scattered atomic-scale defects caused chiefly by background radiation. The materials-, component-, and system-level analyses in Chapters 9-12 have shown how this more precise control can yield systems of high performance. Designs that tolerate the residual flaws caused by radiation damage commonly sacrifice performance, but the lesser size and density of the flaws results in lesser sacrifices. For structures small enough to avoid damage, or large enough to be inherently tolerant of it, the reproducibility of properties and performance can approach perfection. This improves reliability.

b. Strength and stiffness. It is well known that intrinsically strong materials (e.g., ceramics such as aluminum oxide, silicon carbide, and diamond) are brittle, and that brittle materials placed in tension are sensitive to microcracks (Kelly, 1973). Accordingly, the ability of current engineering practice to exploit the high intrinsic strength of these materials is sharply limited by the inability of present manufacturing processes to make them without defects. The ability of molecular manufacturing to build structures with only atomic-scale defects permits these materials to be used at nearly their full theoretical strength. Diamond, in particular, has 55\sim 55 and 75 times the ratio of tensile strength to density of strong steel and aluminum alloys, respectively [assuming a conservative 50GPa50 \mathrm{GPa} tensile strength for diamond, and tabulated values for AISI No. 9255 steel, hardest temper, and 7178-T6 aluminum (Tapley and Poston, 1990)].

The design of minimum-mass compressive structures often is constrained by stiffness, and the ratio of Young's modulus to density for diamond exceeds those for steel and aluminum by factors of 12 and 15 , respectively. The design of compressive structures is often further constrained by buckling, and strategies for preventing buckling are often constrained by the scale of initial imperfections and by the difficulty of fabricating the optimum shapes (which can be intricate, and can pose minimum-gauge problems). Finally, it is worth noting that small moving parts, high power-density actuators, and fast sensing and control systems can be combined to make macroscopic objects that actively resist deformation, thereby simulating an extremely high modulus material within certain bounds of frequency response, length scale, and applied load.

c. Other material properties. Conventional manufacturing processes are constrained to work with materials that result from bulk (or biological) processes. Although it is difficult to characterize the limits of what such processes can make (particularly through organic synthesis), they control comparatively few variables: chiefly the time history of bulk composition, temperature, and pressure, augmented by shear rates, magnetic fields, and electric fields. Molecular manufacturing processes, in contrast, control a comparable number of variables per moiety during the fabrication process. This strongly suggests that the range of feasible materials (and therefore material properties) is broader. In some instances, however, bulk processes can presumably make materials that are at or near the limits of performance in particular respects.

The processes and exemplar architecture described in this chapter assume a high vacuum environment, permitting free use of highly active reagents with no special provision for local enclosures. This limits the output to product structures of high stability, and in particular, of negligible vapor pressure. Most highstrength materials and many polymers and elastomers meet these criteria, but many familiar products do not. Chapters 15 and 16 , however, describe the use of mechanosynthetic processes in solution environments (with suitable constraints on reagents, mechanisms, and so forth). High vacuum is convenient and often desirable, but it should not be taken as a necessary condition for all molecular manufacturing processes; indeed, thorough control of the immediate environment of a reaction site can be provided in a liquid medium, as it is in enzymes. Convergent assembly via blocks of intermediate scale can impose constraints on structure, but the feasibility of joining a pair of blocks to form dense, continuous diamondoid structures (Section 9.7.3) suggests that these constraints are not severe, in practical terms.

d. Size. Manufacturing processes today rarely produce features less than 100 nm100 \mathrm{~nm} in size, and most products are macroscopic, having features on a scale of 0.001 m0.001 \mathrm{~m} or more. Molecular manufacturing can construct objects with 0.3 nm\sim 0.3 \mathrm{~nm} features, placing 108\sim 10^{8} distinct atomic features in a volume of (100 nm)3(100 \mathrm{~nm})^{3}. At the opposite end of the size spectrum, both conventional and molecular manufacturing processes built along the lines of the exemplar system can construct objects of indefinitely large size (in both instances, by assembling macroscopic components to form larger structures).

e. Energy conversion. Section 11.7 describes a class of submicron electrostatic motors capable of converting between mechanical and electrical power in either direction at a power density of 1015 W/m3\sim 10^{15} \mathrm{~W} / \mathrm{m}^{3} and an estimated efficiency >.99>.99. This power density is several orders of magnitude greater than that of electric motors produced by current manufacturing processes.

Section 13.3.8 uses previous calculations regarding mechanosynthetic processes to conclude that chemical and mechanical energy can be interconverted with an efficiency >.99>.99 at a power density of 2×109 W/m3\sim 2 \times 10^{9} \mathrm{~W} / \mathrm{m}^{3}, or 8×106 W/kg\sim 8 \times 10^{6} \mathrm{~W} / \mathrm{kg}. If suitably low-dissipation designs can be found (this chiefly entails avoiding major nonisothermal compression losses), then higher speeds of motion (e.g., 1\sim 1 to 10 m/s10 \mathrm{~m} / \mathrm{s}, rather than 0.004 m/s\sim 0.004 \mathrm{~m} / \mathrm{s} ) will be readily achievable with good efficiency, and the power density values that are achievable will increase to 1012 W/m3\sim 10^{12} \mathrm{~W} / \mathrm{m}^{3}, or 1010 W/kg\sim 10^{10} \mathrm{~W} / \mathrm{kg}.

Conversion of optical to electrical energy has applications in power supply. Molecular manufacturing techniques can be used to make multilayer, multiband-gap photovoltaic cells; structures of this sort can convert solar energy to electrical energy with efficiencies substantially greater than . 3 (Hubbard, 1989).

Conversion of electrical to optical energy has applications in display and illumination. The use of quantum-well nanostructures is already of technological interest for improving the performance of light emitting diodes and solidstate lasers, hence the nanoscale structural control provided by molecular manufacturing techniques should lead to significant improvements. Isotropic lightemitting structures have no minimum scale, but owing to diffraction, good directionality requires a size that is large compared to an optical wavelength.

f. Energy storage. Nearly reversible mechanochemical energy conversion can serve as a basis for energy storage. For example, a device that nearly reversibly transforms butane and oxygen into water and carbon dioxide can store 4×107 J/kg\sim 4 \times 10^{7} \mathrm{~J} / \mathrm{kg}, based on the mass of the initial butane, or 9×106 J/kg\sim 9 \times 10^{6} \mathrm{~J} / \mathrm{kg}, based on the total mass of the reactants. Rates of energy storage and release are as described in Section 13.3.8 (note that the power conversion rates possible in macroscopic systems will commonly be limited by cooling). Systems of this sort have substantially higher energy densities than those of storage cells that can be produced using present manufacturing technologies, and power densities that are orders of magnitude higher.

g. Computational capacity. Molecular manufacturing processes can be used to construct computational systems like those described in Chapter 12 (or better, given the conservatism of the design rules employed there). In terms of component densities, some of the relevant system-level parameters include 1019CPUs/m3\sim 10^{19} \mathrm{CPUs} / \mathrm{m}^{3} (Section 12.7), 1025bits/m3\sim 10^{25} \mathrm{bits} / \mathrm{m}^{3} (for RAM storage, Section 12.6.3), and 1028bits/m3\sim 10^{28} \mathrm{bits} / \mathrm{m}^{3} (for tape storage, Section 12.6.4). Individual CPU speeds can equal or exceed 10910^{9} instructions per second (Section 12.5.2c) with an energy consumption of 1016 J\sim 10^{-16} \mathrm{~J} per instruction (Section 12.7.4). Aside from individual CPU speed, each of these parameters exceeds by several orders of magnitude the performance possible in systems made by conventional manufacturing processes.

14.5.6. Manufacturing costs

a. The feasibility of making cost estimates. Estimating the cost of a conventional manufacturing process is difficult unless the product closely resembles one for which experience exists. Product costs depend on the costs of components which may or may not be in production. The costs of establishing, staffing, and managing a production process-including the controls necessary to ensure adequate product quality - can be hard to estimate. In the aerospace industry, in particular, the costs of testing and documentation for new technologies are huge. Where a technology differs greatly from existing practice, estimates of manufacturing costs have traditionally been little more than guesses.

The relationship between complexity and manufacturing costs highlights the unusual nature of cost estimation in molecular manufacturing processes. In conventional manufacturing processes, costs increase with the complexity of the product being made: more intricate systems require more parts and manufacturing operations. In molecular manufacturing processes, each moiety is treated as a distinct part, regardless of the apparent complexity or simplicity of the product. Accordingly, production cost will be essentially independent of complexity (Section 10.10.2d).

In molecular manufacturing, many of the traditional costs of building complex systems are avoided. Most of the costs that remain are sufficiently well defined that one can place upper bounds on them. These costs are estimated on a per-kilogram basis; for comparison, most manufactured products today fall in the cost range between 10\textdollar/kg\sim 10 \textdollar / \mathrm{kg} (e.g., automobiles, appliances) and 104\textdollar/kg\sim 10^{4} \textdollar / \mathrm{kg} (e.g., aerospace vehicles).

The great exception to this generalization is development costs. These are hard to estimate, but are addressed in Section 14.6, and indirectly by the content of Chapters 15 and 16.

b. Materials. As has been discussed, the chief anticipated feedstock materials for molecular manufacturing (C,N,O,H)(\mathrm{C}, \mathrm{N}, \mathrm{O}, \mathrm{H}) are available in bulk compounds for costs of 0.1\textdollar/kg\sim 0.1 \textdollar / \mathrm{kg} (at a substantial energy cost, they are available from atmospheric gases). Others that may play substantial roles ( Si,P,S,F,Cl\mathrm{Si}, \mathrm{P}, \mathrm{S}, \mathrm{F}, \mathrm{Cl} ) are also available in reasonably inexpensive forms. Precious metals (e.g., of the platinum group) can be expected to be applied as catalysts in mechanochemical processes, but if used at only 10410^{4} cycles per second, a kilogram of platinum atoms (costing 105\textdollar\sim 10^{5} \textdollar ) can be used to process >1010 kg>10^{10} \mathrm{~kg} of material per year, adding little to the product cost.

c. Energy. Using typical organic feedstocks, and assuming oxidation of surplus hydrogen, reasonably efficient molecular manufacturing processes are net energy producers (Section 14.4.8). At a typical price for electrical energy today, 0.1\textdollar/kWhr\sim 0.1 \textdollar / \mathrm{kW} \cdot \mathrm{hr}, the value of the by-product electrical energy would usually exceed the cost of the feedstock materials. The value of by-product energy will be ignored in the present cost estimate.

d. Waste disposal. The only wastes from the manufacturing process itself are high-purity water and a modest amount of heat rejected to the atmosphere (for a 1 kg/hr1 \mathrm{~kg} / \mathrm{hr} system, this is roughly equal to the heat produced by sunlight striking a square meter of dark surface). Neither of these would now be regarded as a significant form of waste. The cost of waste disposal from the preparation of feedstock materials is (within present environmental standards) reflected in the cost of those materials.

e. Labor and distribution. Between the input of raw materials and the removal of finished products, no labor is necessary. Both of these steps can be regarded as distribution costs falling outside the scope of manufacturing proper.

f. Land. The exemplar manufacturing system produces its own mass in product in less than an hour. Conventional manufacturing systems (which must in a fair comparison be taken to include factories that supply parts) take orders of magnitude longer to produce their own mass in product. Assuming comparable ratios of plant and product densities, this indicates that molecular manufacturing systems require orders of magnitude less space (and land) per unit of productive capacity. Costs of land will be reduced by the same factor.

g. Capital. In current manufacturing systems, the cost of capital goods is chiefly a manufacturing cost. In considering a new manufacturing system, this introduces a circular dependency in cost estimation: the cost of producing capital goods depends on the costs of materials, energy, waste disposal, labor, distribution, and capital goods. The equilibrium contribution of the cost of capital goods to the cost of later capital goods can be estimated as follows: Let rr be the real interest rate (expressed as a fractional rate of return per year), cc be the noncapital cost of producing a unit of capital goods, and tt be the time (in years) for a unit of capital goods to produce another unit of capital goods (here treated as a batch process). The total cost CC of a unit of capital goods is then

C=c+C[exp(rt)1]c1rt,rt1\begin{align*} C & =c+C[\exp (r t)-1] \tag{14.3}\\ & \approx \frac{c}{1-r t}, \quad r t \ll 1 \tag{14.4} \end{align*}

In conventional manufacturing, r0.1yr1r \approx 0.1 \mathrm{yr}^{-1}, and t1yrt \approx 1 \mathrm{yr} (or more, particularly if system boundaries are drawn to include factories that supply parts), and the contribution of the cost of capital goods is substantial. In molecular manufacturing, t104yrt \approx 10^{-4} \mathrm{yr}, and these costs are negligible. In this connection, it might be noted that the cost of capital goods contributes to the cost of materials and energy, but the cost benefits of changes in the resulting circular dependencies will not be considered here.

h. Total costs of manufacturing. The preceding discussion indicates that the basic cost of production, here taken to exclude the costs of development and distribution (and of such imponderables as taxation, licensing agreements, and insurance), will be almost wholly determined by the cost of materials. The relevant materials presently cost 0.1\sim 0.1 to 0.5\textdollar/kg0.5 \textdollar / \mathrm{kg}; this range can thus be taken as an upper bound on the basic cost of producing objects using a mature molecular manufacturing technology base. These costs are low enough for molecular manufacturing to be competitive in making a wide range of products.

14.6. Design and complexity

The design of systems that organize complex patterns of activity among reliable, submicron devices operating at millions of cycles per second is an everyday activity in computer science. Although molecular machines and manufacturing differ greatly from microelectronics and computation, many similar issues arise. It is reasonable to expect that computer scientists will play a leading role in the development of molecular manufacturing systems at levels above device design. Accordingly, this section repeatedly compares established concepts in computer science with the anticipated needs of molecular manufacturing.

Device design will likewise depend on computer science. Although experimentation is likely to play a large role, simulations and computer-aided design

(CAD) tools will be essential. Complex systems will become possible only when an adequate library of simple components and operations has been accumulated. The following discussion will assume the previous development of such a library, thereby sharpening the focus on issues arising from complexity.

14.6.1. Part counts and automation in design and computation

Many of the numbers that describe the manufacture of macroscopic objects by molecular means are enormous by most present standards. Building a cubicmicron block requires 1011\sim 10^{11} moietal assembly operations. Building a kilogram object from cubic-micron blocks requires 1015\sim 10^{15} block assembly operations. A team of a thousand people each making one design decision per second would require a century of 8-hour days to plan a single cubic micron, to say nothing of a kilogram: in a practical design process, most of these operations must be specified without direct human attention. Most operations must be as routine as painting a pixel in a raster output device or executing an instruction in a piece of code.

For comparison, typesetting this volume required that a complex pattern of >1010>10^{10} pixels be painted black or white; these operations were directed by a set of single-purpose PostScript programs that were automatically generated from >108>10^{8} bits of data describing text and graphics. The software used to prepare this data executed >1013>10^{13} instructions (though mostly while waiting for inputs). The software itself consisted of >107>10^{7} instructions, written by several teams to form separate programs that nonetheless cooperated to produce a single product. It is reasonable to assume that molecular manufacturing, which likewise generates complex patterns through repetitive operations, will involve broadly similar operations controlled by broadly similar software systems.

14.6.2. Design of components and small systems

Today, most machine code that directs computer operations is generated automatically from a more abstract description in a high-level language. The compiler that does this works by piecing together small, standardized fragments of machine code (consisting of one or more instructions) to implement the desired operations. These fragments and the rules for generating sequences of them are the result of human attention, but the results of this labor are used repeatedly without human attention. In a similar fashion, component-level nanomechanical designs and rules for their composition can be designed with human attention, then later combined by a hardware compiler to implement desired functions.

Examples of direct atom-by-atom design of molecular mechanical devices are scattered through this volume. Some designs, however, show the early results of automation, in which nanometer-scale components are described by languages that specify regular arrays of atoms and bonds. A good example is R. Merkle's prototype computer-aided design (CAD) software that starts with a single typed line of input parameters and generates the atomic coordinates of a regular, three-dimensional diamondoid structure that can contain many thousands of atoms and bonds (Merkle, 1991). Examples of its outputs include the shaft and sleeve in Figure 10.17 and the ring gear in Figure 10.32.

The design of less regular structures can also be automated. A precompiled library of Kaehler brackets (Section 9.5.5) can be used to choose bridging structures between surfaces. Simulated annealing methods (with change of valence and atom type as an operator) hold promise for generating regions of amorphous structure that meet particular boundary conditions.

Eventually, molecular CAD tools will combine principles like those applied in Merkle's and Kaehler's prototypes with features like those found in commercial CAD systems for macroscopic design, such as support for continuum descriptions of objects containing many atoms. With these CAD tools (and faster machines, better molecular potentials, and feedback from physical experiments) it will be possible to design components and systems of the complexity that can be managed today in the macroscopic world. These will include manipulator arms, computers, and molecular mill systems using multiple belts, rollers, and cam surfaces. Components and systems of this complexity can then serve as building blocks in the design of more complex systems.

14.6.3. Automated generation of synthesis and assembly procedures

a. Retrosynthetic analysis of chemical syntheses. Once the structure of a component is specified, a synthesis must be found or the component must be redesigned. Software for the automated generation of candidate chemical syntheses has been in use since the 1960s (see Corey and Cheng, 1989). These programs apply retrosynthetic analysis, the development and application of which won E. J. Corey the 1990 Nobel prize for chemistry. The central idea of retrosynthetic analysis to plan the assembly of a complex molecular structure by considering how it could be disassembled into small, available molecules using imaginary operations, each the reverse of a feasible synthetic step. The resulting disassembly process is therefore the reverse of a feasible synthesis. (Computer scientists will recognize this procedure as a form of backward chaining.)

Mechanosynthetic operations are equally subject to retrosynthetic analysis (but largely free of the complication of competing reactions). Each synthetic step can be regarded as the reverse of an etching reaction, and the synthesis designed (in reverse) as the etching away of the desired structure. Considering the mechanosynthesis of diamondoid structures, Soreff (1992) has suggested that a study of actual etching processes (e.g., of diamond by fluorine) can indicate geometries and transition states for synthetic steps; the forward-direction reagents would be analogous to the hypothetical etching products, but with bond strengths (etc.) favoring deposition over etching.

b. Hierarchical decomposition of larger structures. If an object is to be made from an array of parts, it (and its assembly procedure) can be designed by dividing it into regions, each made of subregions, and so on down to the level of the individual parts. The divisions chosen during the design phase must create interfaces that permit assembly during the manufacturing phase. If the final product were fully specified at the outset, this hierarchical decomposition approach would be a form of retrosynthetic analysis.

In conventional engineering, however, the specification of the final product is influenced by the nature of the available parts and assembly operations. Abstractly, this means that the hierarchical decomposition is developed together with the design, rather than afterward. In dividing a system into regions, the interfaces can be altered as the design becomes more detailed. A design process that includes an assembly-oriented hierarchical decomposition can directly generate the outlines of an assembly plan, and can provide a framework for dividing the computational load of the design process over a large number of processors. (For other purposes, designing in terms of functional rather than geometrical decompositions will be important, requiring multiple representations of the design.) For an early essay on the general applicability of hierarchical decompositions, see Simon (1981).

14.6.4. Shape description languages and part arrays

A shape description language can describe a shape so as to enable an interpreter driving an output device to make a corresponding physical structure. In a program in the PostScript language, which describes two-dimensional shapes, a set of instructions can describe a circle by its radius, color, and border. A language for describing three-dimensional shapes might similarly describe a sphere by its radius, material, and surface structure. With PostScript driving a printer, the circle is approximated by printing an array of dots, fixed with respect to a page; with a three-dimensional language driving a molecular manufacturing system, a sphere would be approximated by assembling an array of blocks, and, unless specified as joined to another object, would remain mobile. In either case, the complexity of the description is independent of the size of the object and hence of the number of pieces from which it is made.

Three-dimensional shape description languages are standards in computeraided design systems for mechanical engineering. An interpreter operating on a CAD description could generate instructions for a molecular manufacturing system, directing it to make a corresponding set of solid objects. Using cubicmicron blocks that can be composed to make solid regions of widely varying properties, the output structures could form a wide variety of high-performance systems with submicron control of component shapes. Not all CAD descriptions will correspond to feasible assembly sequences, but many structures can be made to fall within the required design rules (for example, that each object be supported while it is being built, and that each object be subject to a hierarchical decomposition forming suitable interfaces).

14.6.5. Compilers

a. Assembly-process compilers. Given a fully specified design that meets the constraints necessary to permit assembly, there remains the task of specifying the operations of the assembly process. If the product is to be made from micron-scale blocks, this amounts to specifying the transport of blocks of the correct types to the correct assembly workspaces in the correct order, and specifying the sequence of motions to be executed by manipulators in putting the blocks together. As mentioned in Section 14.6.3b, the design of structures and assembly procedures by hierarchical decomposition directly generates a tree of assembly steps. If this tree has been chosen to generate parts of the appropriate sizes and numbers, then it can be mapped onto a manufacturing layout (perhaps one not entirely unlike that in Figure 14.4). The vast number of manipulator motions to be specified at the finest, earliest, and most-dispersed levels of the assembly process can be planned in parallel by identical software systems running almost independently on separate processors. The result of this parallel assembly-process compilation is a set of instructions that, when executed by the transportation system and manipulator controllers of a manufacturing mechanism, will result in the assembly of an object corresponding to the initial design.

b. Design compilers. A conventional software compiler generates a set of machine instructions from a more abstract specification: a program written in a high-level language. An assembly-process compiler likewise generates a set of machine instructions from an abstract specification, the hierarchically decomposed structure of a particular object (which could itself be specified by a shapedescription language). Like a software compiler, it acts as a compiler with respect to the machine instructions, although starting with a different kind of specification. A modern silicon compiler, in contrast, generates a complex pattern of transistors and conductors from an abstract specification of the properties of a digital circuit. It thus acts as a compiler with respect to the design of a machine. The implementation of artifacts as complex as the exemplar architecture of Section 14.4 seems likely to require the use of analogous design compilers for nanomechanical systems.

There can be no fully general design compilers, capable of converting an arbitrary design specification into a design: not all specifications can be met, and of those that can, not all can be met by the application of compiler-style rules guiding the composition of stereotyped solutions. Like a high-level language compiler or a silicon compiler, a design compiler will accept only specifications for a restricted class of mechanisms in a restricted format.

A compiler for the design of manufacturing systems can only be developed after solutions have been developed to each of the stereotyped, smaller-scale problems that will be encountered. These include:

  • Software modules for performing retrosynthetic analyses of components.
  • Specifications of reaction geometries and environments for effecting standard synthesis steps in mill systems.
  • Designs for subsystems that assemble special components (Section 14.2.1).
  • Designs for structural elements, reagent devices, mill mechanisms, gauges, manipulator arms, computers, communication channels, and so forth.
  • Parameterized descriptions of subsystems for transportation, power distribution, communication networks, and so forth.
  • Design rules for composing subsystems and selecting spatial layouts.

A further requirement, of course, is experimental feedback to permit debugging of system and subsystem designs that fail to work. Note that compiler-aided design need not all be performed by a single tool with seamless end-to-end automation.

The compilation of manufacturing systems will have similarities to the compilation both of digital circuits and of machine code: The product of the compilation process must cause large numbers of operations to be performed in the correct sequence. In designing this product, trade-offs involving time, space, and power must be made, taking into account the relative frequencies and costs of different operations. Preexisting solutions developed for small-scale problems must be described and organized so that they can be applied by algorithms. Throughput, latency, buffers, storage, and transport paths must all be considered. Compact arrangement of parts in space is important (albeit in three dimensions, rather than the two of silicon compilers).

c. The economics of compiler development. Compilers have not been developed for macroscopic mechanical systems, { }^{*} presumably because either (1) the cost-benefit ratio in developing and using them has been adverse, or (2) they are impossible to develop within the constraints of present computer hardware and software technology. If (2) were true because of software technology constraints, this would cast doubt on the feasibility of developing compilers for nanomechanical systems. The history of software and chip-design compilers sheds light on this situation.

When processors were orders of magnitude slower and memory orders of magnitude more expensive, programs were short, and it was crucial to minimize both size and execution time. Small programs could be described in assembly language, providing instruction-by-instruction control of the machine, and there were great incentives to do so. The first computers became operational in the mid-1940s, but the first high-level language to achieve widespread use (Fortran II) was introduced over a decade later, in 1958. Through the 1970s, machine cycles and memory space were scarce enough, and the quality of compiled code poor enough, that assembly language continued to find widespread use in critical programs. Today, faster machines and cheaper memory have enabled the development of programs too large and complex for practical assembly language coding, and improvements in speed, cost, and compilers have combined to make assembly-language coding a shrinking part of software development.

To summarize the abstract lesson: When resources are scarce, systems must be relatively small and simple, and detailed human planning is both feasible and desirable. The incentive to develop compilers is low, and the evolution of compilers good enough to compete with human designers does not even begin. When resources are abundant and inexpensive, detailed human planning cannot organize them on a sufficient scale to exploit their potential. Even inefficient compilers become attractive. They are implemented, and evolve toward greater efficiency.

In macroscopic hardware design today, parts are few and production costs are large. Relatively crude compilers could not compete with human designers in the quality and cost effectiveness of their designs. People can do the job, and do it better. With molecular manufacturing, however, part counts can grow into the trillions and beyond, and per-part costs will fall to minuscule fractions of a dollar. Limiting designs to the number of parts that an unaided team of human designers can manage will be unattractive, if a compiler-aided design team can specify a system a million times more complex and capable. This preference would often remain true even if each compiler-specified system wastes twice as much space, mass, and energy as would a comparable system specified in detail by a competent human designer.3

In a design process following the present compiler model, human design will remain dominant at the level of parts and subsystems (in the form of knowledge built into the compiler) and at the level of overall system organization and purpose (in the form of specifications given to the compiler when it is used). The intermediate levels will be designed, with considerable inefficiency, using algorithms and heuristics that represent a workable subset of human knowledge of design principles.

This pattern is, in the early 1990s, already familiar in one hardware domain. Since the early 1960s, the number of components in an integrated circuit has climbed by a factor of 106\sim 10^{6}, and the manufacturing cost per transistor has fallen to 105\sim 10^{-5} dollars or less. If a 10610^{6} transistor design has an expected market of 10510^{5} units, then every dollar of design cost per transistor adds ten dollars to the price of each chip, yet a dollar cannot buy much time from a human design team. The design cost is often a large fraction of the product cost. In response to this pressure, silicon compiler technology emerged in the 1980s. At first considerably less efficient than a human designer in maximizing speed and minimizing circuit size and cost, silicon compilers nonetheless gained a foothold, then steadily improved, becoming an integral part of the design process.

This section has given only a rough sketch of past developments. A more detailed study of the evolution from lower-level CAD systems to silicon compilers could yield further insights into the likely path from today's molecular modeling and mechanical CAD packages toward future compiler-style support for the engineering of nanomechanical systems.

14.6.6. Relative complexities

As suggested by the discussion in Section 14.6.4, many nanomechanical systems are relatively simple. Highly regular structures provide an obvious set of examples. Among more complex structures, mechanical nanocomputers are of essentially the same complexity as microelectronic computers, once a modest library of logic-element designs has been accumulated (indeed, some CAD software for digital logic design would be applicable without modification). Regular arrays of nanocomputers with local interconnection will not be substantially more complex than single computers, hence the design of extremely complex systems can eventually rely on such arrays to solve computationally expensive problems.

Small molecular manufacturing systems are comparable in complexity to modern automated factories, both in parts count and in organization. Since macroscopic automated factories fall within the range of present human design capabilities, their nanomechanical equivalents should as well. The focus in this chapter on large systems of great complexity should not be taken to imply that all useful manufacturing systems need be so complex.

Even the most complex systems described here are arguably less complex than some modern software systems. Although their part counts greatly exceed the number of lines of code in any program, parts and lines of code are not analogous. Mechanisms in a manufacturing system will appear in many copies; a particular mechanism (and its parts) is less like a unique subroutine or object class (and its code) than like one of many invocations of that subroutine, or instances of that object. Indeed, it is often like an invocation or instance with variables identical to those already used, existing in multiple copies in order to produce identical "constants," or building blocks, for repeated use in a physical product.4 Writing a program with 101510^{15} lines of code would be out of the question today on grounds of complexity. Running a program that executes 101510^{15} instructions is less extraordinary, amounting to a month of supercomputer time. The difference between these cases lies in the massive repetition of standard operations. Nanomechanical systems can exhibit a similarly massive repetition of physical structures, without necessarily entailing unprecedented complexity.

14.7. Conclusions

Molecular mill and manipulator mechanisms can be combined in architectures capable of transforming solutions of small organic molecules into complex macroscopic artifacts built of diamondoid materials and containing components of the sorts described in Chapters 9-13. A 1 kg1 \mathrm{~kg} system can build a 1 kg1 \mathrm{~kg} product object in 1hr\sim 1 \mathrm{hr}, dissipating 1.1 kW\sim 1.1 \mathrm{~kW} of heat. The range of products that can be produced is large, encompassing high-performance structures, massively parallel supercomputers, and additional molecular manufacturing systems. A review of costs (other than those of design, licensing, taxation, and the like) indicates that the cost of producing objects using such systems can approach the cost of the required feedstocks, for exoergic reactions. Assuming present materials prices, the anticipated product costs would be 0.10\sim 0.10 to 0.50\textdollar/kg0.50 \textdollar / \mathrm{kg}. The complexity of macroscopic molecular manufacturing systems will require extensive automation of the design process.

Some open problems. Far more detailed system-level descriptions of molecular manufacturing could be provided by studies that first attempt to clarify the relationship between (1) the number of distinct building-block types and (2) the range of capabilities that can be delivered by products made from those building blocks. Alternative architectures are worth considering, including architectures that incorporate greater flexibility and more extensive use of generalpurpose mechanisms. Section 14.6 can be regarded as a partial survey of open problems in automating the design of products and the specification of production processes. The literature on analogous macroscale problems is a natural point of departure for work in this area.


  1. Damage to a large fraction of units in a set increases the operation and damage rates of the remaining units, but this makes a negligible adverse contribution to overall system reliability.

  2. P. Barth has expressed concern that breaching the vacuum containment of a manufacturing system would cause its premature oxidation (i.e., that the finely-divided, carbonrich nanomechanical components would burn). Note, however, that little of the mass of the system described serves a chemical or interfacial role; most of the mass serves as structure (including the interiors of moving parts). In systems to be used in environments where combustibility is a safety concern, much of this structural mass can be made from incombustible materials (e.g., aluminum oxide; see Table 9.1) with little change in overall design. Internal water reservoirs may also be useful.

  3. CAD systems are automating more functions, and research in entirely automated design is underway. Accordingly, this statement is becoming debatable.

  4. The computational analogue would be a comparatively simple massively parallel SISD machine, pointlessly producing many instances of the same answer. In manufacturing, applying the same operation ("single instruction") to identical objects ("single datum") is more useful, producing multiple instances of a physical object.