Self replicating systems and low cost manufacturing

Ralph C. Merkle
Xerox PARC
3333 Coyote Hill Road
Palo Alto, CA 94304
merkle@xerox.com

Further information on self replicating systems is available at http://www.zyvex.com/nanotech/selfRep.html

This paper was first published in The Ultimate Limits of Fabrication and Measurement, M.E. Welland, J.K. Gimzewski, eds.; Kluwer, Dordrecht, 1994, pages 25-32.
This electronic reprint is available on the web at http://www.zyvex.com/nanotech/selfRepNATO.html, and might differ from the printed version.

1. Introduction

Experimental[7] and theoretical[4, 6, 8, 21, 22] work both support the idea that we will be able to fabricate precise molecular structures (such as molecular logic elements) by positioning individual atoms and molecules. However, even the ability to make and interconnect a few atomically precise logic elements will have limited impact when we must make and interconnect at least trillions of logic elements to surpass projected future lithographic capabilities.

The only demonstrated method of mass producing complex highly precise structures at a low cost per kilogram is by programmable self replicating systems as exemplified by potatoes, wheat, wood, etc. (Electronics are not cheap: on a per kilogram basis they are more than one hundred times as expensive as gold). Unfortunately, it's not clear that such biological methods will be able to produce the full range of products we desire. Many of today's products are not made of biological material and there is no particular reason to believe this situation will change. Today's artificial computers are not made out of protein because other materials offer superior performance. Biological computers, despite their many virtues, have high error rates, millisecond logic delays and meter-per-second signal propagation speeds: they are grossly uncompetitive.

While the design and development of non-biological programmable self replicating systems suited to the manufacture of complex high performance computer systems (as well as a range of other high precision products) might at first appear daunting, there has been much theoretical work in this area. Starting with von Neumann's "universal constructor" and "kinematic machine" in the 1950's and continuing through the more recent proposals by Drexler for an "assembler" this work describes a range of possible system designs. Many of these systems are not overly complex by today's engineering standards. More recent work suggests that further simplifications are possible and that research to determine the simplest and most easily manufacturable programmable self replicating system should be pursued.

2. General manufacturing systems

Because biological self replicating systems are so ubiquitous it is common to assume that their specific properties and idiosyncratic features are an inherent requirement for all self replicating systems. However, programmable self replicating systems designed for manufacturing need bear little resemblance to biological systems. We shall call such non-biological systems general manufacturing systems. In this article we highlight the differences between biological systems and general manufacturing systems.

Design concepts for general manufacturing systems have been discussed for many years [10, 27, 28], and their utility in manufacturing has been emphasized recently [4, 5, 6, 18]. These proposals draw on a body of work started by von Neumann[27]. A wide range of methods have been considered[10, particularly pages 190 et sequitur "Theoretical Background"]. The von Neumann architecture for a self replicating system is the ancestral and archetypal proposal[24, 27].

2. The von Neumann architecture for a general manufacturing system

Von Neumann's proposal consisted of two central elements: a universal computer and a universal constructor (see figure 1). The universal computer contains a program that directs the behavior of the universal constructor. The universal constructor, in turn, is used to manufacture both another universal computer and another universal constructor. Once construction is finished the program contained in the original universal computer is copied to the new universal computer and program execution is started.

Von Neumann worked out the details for a constructor that worked in a theoretical two-dimensional cellular automata world (parts of his proposal have since been modeled computationally[24]). The constructor had an arm which it could move about and which could be used to change the state of the cell at the tip of the arm. By progressively sweeping the arm back and forth and changing the state of the cell at the tip, it was possible to create "objects" consisting of regions of the two-dimensional cellular automata world which were fully specified by the program that controlled the constructor.

While this solution demonstrates the theoretical validity of the idea, von Neumann's kinematic constructor (which was not worked out in such detail) has had perhaps a greater influence, for it is a model of general manufacturing which can more easily be adapted to the three-dimensional world in which we live. The kinematic constructor was a robotic arm which moved in three-space and which grasped parts from a sea of parts around it. These parts were then assembled into another kinematic constructor and its associated control computer.

An important point to notice is that self replication, while important, is not by itself an objective. A device able to make copies of itself but unable to make anything else would not be very valuable. Von Neumann's proposals centered around the combination of a universal constructor, which could make anything it was directed to make, and a universal computer, which could compute anything it was directed to compute. It is this ability to make any one of a broad range of structures under flexible programmatic control that is of value. The ability of the device to make copies of itself is simply a means to achieve low cost, rather than an end in itself.

3. Drexler's architecture for an assembler

Drexler's assembler follows the von Neumann kinematic architecture, but is specialized for dealing with systems made of atoms. The essential components in Drexler's assembler are shown in figure 2. The emphasis here (in contrast to von Neumann's proposal) is on small size. The computer and constructor both shrink to the molecular scale, while the constructor takes on additional detail consistent with the desire to manipulate molecular structures with atomic precision. The molecular constructor has two major subsystems: (1) a positional capability and (2) the tip chemistry.

The positional capability might be provided by one or more small robotic arms, or alternatively might be provided by any one of a wide range of devices that provide positional control[9, 15]. The emphasis, though, is on a positional device that is very small in scale: perhaps 0.1 microns (100 nanometers) or so in size.

The tip chemistry is logically similar to the ability of the von Neumann universal constructor to alter the state of a cell at the tip of the arm, but now the change in "state" corresponds to a change in molecular structure. That is, we must specify a set of well defined chemical reactions that take place at the tip of the arm, and this set must be sufficient to allow the synthesis of the structures of interest.

It is worth noting that current methods in computational chemistry are sufficient to model the kinds of structures that will appear in a broad class of molecular machines, including all of the structures and reactions needed for some assemblers[16, 20, 21, 22]

4. The Broadcast Architecture

In the von Neumann architecture, Drexler's assembler and in living systems the complete set of plans for the system are carried internally in some sort of memory. This is not a logical necessity in a general manufacturing system. If we separate the "constructor" from the "computer," and allow many individual constructors to receive broadcast instructions from a single central computer then each constructor need not remember the plans for what it is going to construct: it can simply be told what to do as it does it (see figure 3). This approach not only eliminates the requirement for a central repository of plans within the constructor (which is now the component that self replicates), it can also eliminate almost all of the mechanisms involved in decoding and interpreting those plans. The advantages of the broadcast architecture are: (1) it reduces the size and complexity of the self replicating component, (2) it allows the self replicating component to be rapidly redirected to build something novel, and (3) If the central computer is macroscopic and under our direct control, the broadcast architecture is inherently safe in that the individual constructors lack sufficient capability to function autonomously[6, 18].

This general approach is similar to that taken in the Connection Machine[14], in which a single complex central processor broadcasts instructions to a large number of very simple processors. Storing the program, decoding instructions, and other common activities are the responsibility of the single central processor; while the large number of small processors need only interpret a small set of very simple instructions.

It is interesting to view the cell as using the broadcast architecture with the nucleus as the "central computer" broadcasting instructions in the form of mRNA to perhaps millions[29] of ribosomes.

Drexler has proposed immersing the constructor in a liquid or gas capable of transmitting pressure changes and using pressure sensitive ratchets to control the motions of the constructor[6]. If each pressure sensitive ratchet has a distinct pressure threshold (so that pressure transitions around the threshold cause the ratchet to cycle through a sequence of steps while pressure changes that remain above or below the threshold cause the ratchet to remain inoperative) then it is possible to address individual ratchets simply by adjusting the pressure of the surrounding fluid. This greatly reduces the complexity of the instruction decoding hardware.

5. Differences between biological systems and general manufacturing systems

General manufacturing systems are likely to be very different from biological systems. First, general manufacturing systems aim to produce products with the best achievable performance and capabilities, e.g., which approach the fundamental limits imposed by physics and chemistry. Biological systems, based largely on protein, are unlikely to achieve this objective. Second, it seems likely that the indirect and circuitous routes by which biological systems control three dimensional structure (e.g., the protein folding problem, self assembly to control the position of molecular components, etc.) will be largely replaced by simpler and more direct methods that use positional control. Third, the error rates in biological systems are relatively high. It should be feasible to substantially reduce these error rates and produce systems and products with superior reliability, performance, materials properties, etc. Fourth, biological systems are not designed to allow rapid reprogramming. A potato cannot readily be reprogrammed to make a steak. General manufacturing systems should be able to respond rapidly to changing requirements by changing what is manufactured. Fifth and last (at least in this paper), we want general manufacturing systems to be free of extraordinary risks.

6. More than proteins

The greater the diversity of products a manufacturing system can make, the more valuable it is. If it can only make biological products, its value is reduced. Consider the problem of building high performance computers. While biological computers (e.g., the human brain and nervous system) have many fine properties (and utilize an architecture and software which is clearly greatly superior in many respects to anything currently available), they are based on fundamental components (synapses, neurons) which have truly atrocious performance. Logic elements with millisecond delays and meter-per-second signal propagation velocities are grossly unacceptable in today's computers, much less in future systems. (Note that the poor performance of the underlying hardware increases our respect for an architecture and software which manage to wring such amazing feats from such slow and unreliable components).

It seems certain that future computers will have the smallest possible logic elements, built with the highest possible precision and at the lowest possible cost. This should result in logic elements which are molecular in both size and precision, assembled in complex and idiosyncratic patterns.

A more plausible candidate than proteins for future computational hardware is semiconductor devices conceptually similar to today's but made with vastly greater precision (individual dopant atoms placed deliberately at specific lattice sites, for example) and which extend fully into three dimensions. Diamond, with its wide band gap, excellent thermal conductivity, large breakdown field and high mobility would provide an excellent semiconductor for such future devices[12]. Molecular-sized logic elements packed densely in three dimensions will produce significant heat; an often overlooked problem in molecular logic proposals. This problem can be dealt with by using thermodynamically reversible logic[19 and references therein].

Biological structural materials are also far from ideal. Diamond has a strength to weight ratio over 50 times that of steel, and properly engineered materials in the future should be able to approach this strength and yet resist fracturing. Nothing in biology approaches this.

The chemical reactions involved in the synthesis of diamond today are very different from those involved in making proteins[1, 2, 11]. Reactions proposed for the atomically precise synthesis of diamondoid structures involve highly reactive compounds in an inert environment[6, 21, 22]; a very different approach than that taken in biological systems. For strength and stiffness, materials using boron, carbon and nitrogen are superior[3]. Diamond is also an excellent candidate material for future electronic devices.

If we limit general manufacturing systems to proteins we will exclude a vast range of very valuable products. We will almost certainly wish to make diamond and diamondoid products. This implies the use of reactions and conditions very different from what we see in biology today.

7. Positional Control

Besides using non-biological materials, general manufacturing systems are likely to make extensive use of positional control, i.e., the ability to position molecular components appropriately by using molecular positional devices. The Stewart platform[9, 13, 25, 26] seems ideal for providing positional control at the molecular level. The basic Stewart platform is an octahedral structure in which one triangular face is designated the "base," the opposing face is designated the "platform," and six adjustable-length struts (which lie along the six edges of the octahedron which are between the base and platform) control the position of the platform. Within an allowed range of motion it provides complete control over the position and orientation of the platform with respect to the base; it provides high stiffness (critical to positional control at the molecular scale); all struts are either in pure tension or pure compression; and it is a simple design. This simplicity suggests that it might be feasible to self-assemble a Stewart platform (e.g., self assemble an octahedron in which the lengths of the struts can be controlled: either statically at the time of self assembly or dynamically in response to an external signal).

The use of positional control in general manufacturing systems is consistent both with the tradition of kinematic devices seen in theoretical proposals[10, 27], with experience from today's macroscopic manufacturing[23], and with theoretical proposals for molecular manufacturing[6, 21].

While biological systems make extensive use of self assembly at the molecular level, positional control is dominant in today's factories (although vibratory bowl feeders[23] are in essence the macroscopic application of principles more commonly associated with self assembly in the face of thermal noise at the molecular level). The application of positional control at the molecular level appears feasible both theoretically and experimentally, and offers striking advantages in the manufacturing process. The reader is invited to consider the difficulties involved in manufacturing a car if positional control were prohibited in the manufacturing process. We can reasonably expect that the application of positional control to molecular synthesis will greatly extend the range of things that can be made[21]. It will also result in artificial systems that are very different from the biological systems with which we are familiar.

8. Reduced error rates

Another likely difference is in the error rates tolerated during assembly. The achievable error rate limits the range of options that can be pursued and in particular limits the feasible module size. (A "module" is here viewed as an assemblage which has a relatively high probability of being manufactured correctly and of functioning correctly, and hence can be discarded in its entirety if there is any failure anywhere within the module). When error rates are high, the module size must be small. If the module size were large in the face of high error rates, the yield of correctly working modules would be unacceptably low and overall system function would be compromised. When error rates are low, the module size can be large. Protein synthesis has an error rate of roughly 1 in 10,000 [29] and we do not find proteins with tens of thousands of amino acids. "Typical" proteins have hundreds or perhaps a few thousand amino acids.

There are well known methods of assembling unreliable logic elements into reliable computational systems. However, these methods result in reduced system performance and increased bulk. Experience with semiconductor devices supports the idea that the primary objective in the manufacturing process is to reduce the error rate to the lowest possible level, and only when further reductions are infeasible should redundant logic elements (or other error- tolerant design approaches) be adopted.

Applying this philosophy to general manufacturing systems, we should first determine the lowest achievable error rate and then design modules of the largest possible size using the simplest and most efficient designs. It seems difficult to reduce error rates at the molecular level substantially below the levels caused by radiation[6]. Other error mechanisms (e.g., thermal, photochemical) can be reduced to levels that are below the error rate caused by radiation damage[6] by using appropriate designs. This conclusion leads to feasible molecular module sizes of tens of billions of atoms with MTBF's of many decades (where an "error" is defined to occur if even a single atom is out of place). This is in sharp contrast to the error rates and module sizes adopted in biological systems. We can reasonably expect that systems that take advantage of these low errors rates will involve designs and system functions that are very different from biological systems.

9. Ease of reprogramming

General manufacturing systems should be so designed that they can readily change what they are manufacturing. While spraying mRNA over plants to cause the rapid manufacture of the desired protein has been proposed, biological systems by and large lack the ability to accept external instructions about what is to be built. In general manufacturing systems, by contrast, we will wish to be able to redirect the manufacturing process quickly and rapidly in response to changing demand.

10. Risks of self replicating systems

Self replicating systems, like other systems, might fail to work correctly and as a consequence cause damage. Unlike ordinary systems, they can theoretically inflict an unlimited amount of damage. They could theoretically, for example, replicate unchecked and destroy the planet[5]. To be acceptable, any proposed general manufacturing system must be inherently safe; i.e., not only must the system as designed not pose any extraordinary risks, this property must be retained even in the face of accidental design errors, errors in handling or transmitting the instructions, etc. It must be robustly safe. There are reasons for believing that general manufacturing systems will lack the marvelous flexibility and adaptability that is characteristic of living organisms and will suffer from the same rigid and inflexible responses to even small changes in the environment that are so common in other machinery[17]. This inflexibility is economically beneficial for it simplifies design and increases efficiency and economy (flexible systems able to adapt to a wide range of environments are imperfectly adapted and less efficient in any specific environment than less flexible systems narrowly tuned to that particular environment). Inflexibility is also desirable as a safety feature, for inflexible systems will fail in an uncontrolled environment.

11. Conclusion

General purpose programmable self replicating systems designed for manufacturing are likely to differ dramatically from biological self replicating systems. Both ducks and 747's fly, but they are very different. Some of the likely differences: (1) general manufacturing systems will employ non-biological reactions to make products of greater strength and superior electronic performance (diamondoid structures being a primary candidate in both cases). (2) They will take full advantage of the principles of positional control as exemplified by devices such as robotic arms, Stewart platforms and the like. (3) They will use basic operations that have a much higher reliability than those used in biological systems, and so will be able to assemble larger modules (more atoms) with good confidence that they are error free. This will allow the exploitation of module designs that are more efficient and compact than anything that could be contemplated with the relatively high error rates seen in biological systems. (4) They will be readily reprogrammable. (5) When designed for manufacturing (and not deliberately designed to be dangerous, as in weapons) they will be unable to replicate outside of a very specific and unnatural environment, making them inherently safe.

12. References

1. Angus, J.C., Argoitia, A., Gat, R., Li, Z., Sunkara, M., Wang, L. and Wang Y. (1993) Chemical vapour deposition of diamond, Phil. Trans. R. Soc. Lond. A, 342, pp. 195-208.
2. Butler, J.E., and Woodin, R. (1993) Thin film diamond growth mechanisms, Phil. Trans. R. Soc. Lond. A, 342 pp. 209-224.
3. Cohen, M.L. (1993) Predicting Useful Materials, Science 261, pp. 307-308.
4. Drexler, K.E. (1981) Molecular Engineering: an approach to the development of general capabilities for molecular manipulation, Proc. National Academy of Sciences USA, 78, pp. 5275-8.
5. Drexler, K. E. (1986) Engines of Creation, Doubleday.
6. Drexler, K.E., (1992) Nanosystems: molecular machinery, manufacturing, and computation, Wiley&Sons.
7. Eigler, D.M., and Schweizer, E.K. (1990) Positioning single atoms with a scanning tunnelling microscope, Nature 344, 524-526.
8. Feynman, R.P. (1960) There's Plenty of Room at the Bottom, Caltech's Engineering and Science, February.
9. Fitzgerald, J.M. and Lewis, F.L. (1993) Evaluating the Stewart Platform for Manufacturing, Robotics Today, 6, pp. 1-3.
10. Freitas, R.A., and Gilbreath, W.P., (1980) Advanced Automation for Space Missions, National Technical Information Service N83-15348.
11. Frenklach, M., and Wang, H. (1991-I) Detailed surface and gas-phase chemical kinetics of diamond deposition, Physical Review B, 43, pp. 1520-1545.
12. Geis, M.W., and Angus, J.C., (1992) Diamond Film Semiconductors, Scientific American, October pp. 84.
13. Gough, V.E., and Whitehall, S.G. (1962) Universal Tyre Test Machine, Proc. 9th Int. Tech Congress F.I.S.I.T.A., pp. 117.
14. Hillis, D. (1986) The Connection Machine, MIT Press.
15. Klafter, R.D., Chmielewski, T.A., and Negin M. (1989) Robotic Engineering: an Integrated Approach, Prentice Hall.
16. Merkle, R.C. (1991) Computational Nanotechnology, Nanotechnology, 2, pp. 134-141.
17. Merkle, R.C. (1992) Risk Assessment, in Crandall, B.C. and Lewis, J. (eds) Nanotechnology: Research and Perspectives, MIT press, pp. 287-294
18. Merkle, R.C. (1992) Self replicating systems and molecular manufacturing, Journal of the British Interplanetary Society, 45, pp. 407-413.
19. Merkle, R.C. (1993) Reversible Electronic Logic Using Switches, Nanotechnology, 4, pp. 21-40.
20. Merkle, R.C. (1993) A Proof About Molecular Bearings, Nanotechnology, 4 pp. 86-90.
21. Merkle, R.C. (1993) Molecular Manufacturing: Adding Positional Control to Chemical Synthesis, Chemical Design Automation News, 8, No. 9&10, pp. 1.
22. Musgrave, C.B., Perry, J.K., Merkle, R.C., and Goddard, W.A. (1992) Theoretical Studies of a Hydrogen Abstraction Tool for Nanotechnology, Nanotechnology, 2, pp. 187-195.
23. Riley, F.J. (1983) Assembly Automation, Industrial Press.
24. Signorini, J. (1989) How a SIMD machine can implement a complex cellular automaton? [sic] A case study: von Neumann's 29-state cellular automaton, in Proceedings Supercomputing `89, ACM Press.
25. Stewart, D (1965-66) A Platform with Six Degrees of Freedom, The Institution of Mechanical Engineers, Proceedings 1965-66, 180 Part 1, No. 15, pp. 371- 386.
26. Stoughton, R.S., and Arai, T. (1993) A Modified Stewart Platform Manipulator with Improved Dexterity, IEEE Transactions on Robotics and Automation, 9, pp. 166-173.
27. von Neumann, J. and Burks, A.W. (1966) Theory of Self-Reproducing Automata, University of Illinois Press.
28. von Tiesenhausen, G., and Darbro, W.A. (1980) Self- Replicating Systems - A Systems Engineering Approach, NASA technical memorandum TM-78304, Marshall Space Flight Center, Alabama.
29. Watson, J.D., Hopkins, N.H., Roberts, J.W., Steitz, J. A., Weiner, A.M. (1987) Molecular Biology of the Gene, Fourth Edition, Benjamin/Cummings.

This page is part of the nanotechnology web site.