If you had $50 M to spend to advance progress on molecular machines, what would you spend it on?

I would spend the $50 million assembling a distributed molecular machine research program composed of multiple teams.

High-Level Budget Breakdown

$25 Million: Theoretical and Computational Research Grants
$25 Million: Experimental Research Grants

On Theoretical and Computational Research

The $25 million in grants would, over a period of 5 years, support 5 teams. Each team would pursue a distinct subfield of molecular machine research that addresses an outstanding question in molecular machine development. Candidate questions could include:

• Risk reduction for self-assembly and replication
• Systemic integration of nanoparts
• Massively parallel mechanosynthesis
• Molecular machine programmability and communication

Basic Science and Engineering Studies

Theoretical research in molecular machines could involve either:

• Basic science for developing models that describe or predict nano-scale behavior, or;
• Engineering research for identifying pathways for yielding functionalities that satisfy technological desiderata.

These two approaches are distinct in their objectives and methods, but can nonetheless complement one another. (Drexler 2004, 3)

Exploratory Engineering and Amenability to Experimentation

Successful research will, for each subfield of molecular machine research, yield a subfield roadmap that refines and articulates a pathway with specified design objectives by which a track of molecular machine development can be defined. The subfield roadmaps should help to foster research and design strategies for redressing key long-term challenges in molecular machine development. The stipends should allow researchers the latitude to study advanced molecular machine capabilities to a degree often inhibited by the pressures of specialization and short-termism imposed by research funders. The objective of the research would be the advancement of “exploratory engineering” studies in molecular machine design that extend to long-term aspirations. (See Drexler 1988) The roadmaps would build upon the 2007 roadmap (see Drexler et al. 2007) but differ therefrom in the following ways:

• Each roadmap should pertain to a particular question in great detail, rather than the entire nanosystems field
• The roadmaps should prioritize basic science and modeling over discussion of application and “policy-oriented” recommendations

Importantly, the short-term spectrum of the roadmap should be compatible with contemporary experimental capabilities such that the relevant designs are amenable to experimental testing.

On Experimental Research

After the second year, two of the research teams with results most amenable to experimentation would be invited to augment their team to include researchers with experimental expertise and facilities in order to subject the short-term specifications of the roadmap to experimental testing. These teams could revise their theoretical and computational models during the fourth year and conduct a second wave of experimentation during the fifth year.

What are design criteria for molecular machines that would enable a transformative application long-term, i.e. 30 years from now?

The literature on molecular machines introduces a number of important criteria that correspond to theoretical capabilities with the potential to generate transformative outcomes. However, it is common for a given paper to be devoted to a singular criterion. Although many proposed capabilities have been theoretically introduced, few papers compare capabilities. Even fewer analyze their compatibility and systemic integrability. Subsuming many capabilities under core criteria, I would nominate the following as the most crucial:

1. Machine reliability (can we trust the machines to perform consistently?)
2. Nanofabrication/manufacturing generalizability (if we design machines, can we build them?)
3. Computational capability (can the machines process information?)

For each of these criteria, a number of capabilities and performance measures can be identified in the literature. The relative advantages of such capabilities and performance measures, as well as potential tradeoffs therebetween, should be better ascertained. A non-exhaustive list of capabilities and performance measures are included below:

Machine Reliability

• Measures:
  • Tensile strength (e.g. Drexler 1995, 325; Freitas and Merkle 2008)
  • Stiffness (e.g. Drexler 1992, 222-224; Drexler and Pamlin 2013, 76)
  • Power density (e.g. Drexler 2005, 341; 2006, 8)
  • Strength/weight ratio (e.g. Freitas 2011, 395)
  • Fault-tolerance and redundancy (Ahuja et al. 2009)
  • Resistance to side reactions (e.g. Freitas 2011, 389)
  • Minimal torsional strain (e.g. Drexler 1996, 396)
  • Productivity (e.g. Freitas 2006)

Fabrication and Manufacturing

• Positional Assembly Capabilities:
  • Design of molecular building blocks (Merkle 2000)
  • Molecular tip arrays (Drexler 1991a, 1991b)
  • Positional control (Merkle 1997a)
• Mechanosynthesis Capabilities:
  • Hydrogen abstraction tool design (e.g. Musgrave et al. 1991)
  • Specification of minimal toolkits (e.g. Freitas and Merkle 2008)
• Self-Assembly Capabilities:
  • Use of assembly stations (Skidmore et al. 2001)
  • Comparison of nanofactory assembly and self-assembly (Phoenix and Drexler 2004)
• Scalability measure:
  • Bottom-up manufacturing through “convergent assembly” (Merkle 1997b)
  • Testing the “Efficient Replicator Scaling Conjecture” (Freitas and Merkle 2004)
• Performance measure:
  • Throughput capability (Drexler et al. 2007)

Nanocomputational Capability:

• Logic Operation Capabilities:
  • Mechanical rod logic (Drexler 1989; 1991, 391-409; 1992, 343-355)
  • Reversible electrode logic (Storrs Hall 1992; 1994a; 1994b, 160-165)
  • Helical logic (Merkle and Drexler 1996)
  • Combinatorial and sequential logic using link-rotary mechanical computation (Merkle et al. 2016, 3-18)

• Hardware Capabilities:
  • Molecular compilers (Merkle 1991, 138)
  • CPU (Drexler 1992, 367-37)
  • Installation of nanocomputers in molecular machines (e.g. Freitas 1998, 419)
• Communication measures:
  • Transmission of instructions from computer to molecular machines (Merkle 2000)
  • Tradeoff between broadcast dependency and machine memory (Merkle 1992)

Commercial application/VC zone: is there any work relevant for progress on molecular machines that is in the zone of “not done, needs to be done, and we have the tools to do”?

The development of a sizable inventory of nanoparts and molecular machine devices (differential and universal gears, bearings, pumps, tooltips, etc.), first beginning in the late 1980s (see Drexler 1987), has been underway for over three decades. And yet, it still remains unclear to what extent compatibilities permit interoperationality between, and systemic integration of, such devices. What has not yet been done, and needs to be done, is the completion of systems design studies for integrating nanoparts. “Larger systems call for multi-scale modeling techniques in which atomistic descriptions support continuum and lumped-component models that describe extended structures and the interaction of subsystems.” (Drexler 2006, 10) The design of larger systems will also require, in turn, the design of larger components: “Atomically precise nanoparts, once fabricated, must be transferred from the fabrication site and assembled into atomically precise complex components containing many nanoparts. Such components may include gear trains in housings, sensors, motors, manipulator arms, power generators, and computers.” (Freitas 2011, 395)

What is your definition of a Molecular Machine? Describe your favorite example with its defining properties

Alternatively, a molecular machine can be defined systemically as a constituent part of a “machine phase system”, a propositional phase of matter in which the components are “eutactic”, or finely arranged. (See Drexler 1992, 6; 2001, 74)

Which enabling technologies, tools, or techniques would you like to have that other fields could provide?

High performance computing can greatly assist the exploration of large configuration spaces in designing molecular machines. Although computational costs are rarely reported in papers on theoretical molecular nanotechnology, they are indeed occasionally reported. In an important paper published by Freitas and Merkle on tooltip design, it is reported that over 100,000 hours of computation was required to conduct the research. (See Freitas and Merkle 2008, 768) Although I am hesitant to extrapolate from this single paper, I do suspect that, if over 100,000 hours of computation is required for tooltip design, simulations of subsystems may require high-performance computing.

What does the field of molecular machines focus on too much/ not enough?

Not Enough: Theoretical Nanotechnology

Too Much: Biochemistry

Admittedly, the study of biological molecular machines is inspirational, edifying, and instructive. However, undue focus on such molecular machines may bias molecular machine researchers towards emulation of such machines or the design of homological devices. The literature on theoretical molecular machines suggests that mechanosynthesis, nanofabrication, and nanocomputation can yield machines without precedent in contemporary biology: “Nature exhibits prototypes of simple productive nanosystems, but physical analysis indicates that artificial systems can go far beyond the biological model.” (Drexler 2005, 339)

What is the best question to ask to drive long-term progress on molecular machines?

In my view, the best question that can be asked is:

“How can researchers contribute towards a generalizable molecular machine capability?”

The question of generalizable research is key because it anchors the field of molecular machine research with reference to the foundational engineering objective that first motivated the field. We can recall that molecular machines were first described in Proceedings of the National Academy of Sciences as “a general molecular engineering technology.” (Drexler 1981, 5275) The notion of generalizability is ideally resistant to departmental/disciplinary specialization and narrow tasks. Instead, a generalizable capability assists the entirety of the molecular machine field.

Some additional remarks can be made on generalizability. An important analogue is the research question of generalizability in artificial intelligence (AI) studies. Researchers pursuing artificial general intelligence (AGI) attempt to develop theories and models of generalizable intelligence with capabilities intended to greatly exceed contemporary ‘narrow AI’ services. (e.g. Goertzel and Pennachin 2007, 1)

I would contend that molecular machine research requires its own analogue of AGI (perhaps, ‘General Molecular Machinery’, or ‘GMM’). A researcher contributing to GMM might produce knowledge, either in basic science or engineering, that fosters an understanding of molecular dynamics and design prospects to which many researchers can refer. Alternatively, a GMM researcher might help to develop a fabrication or computational capability that can be applied to the design of many machines.

In this complex field, what are techniques that can be applied to molecular machine design or production that is actually simple? Chemistry, production techniques, biological hacks, _______, ?

Mathematical modeling techniques, including those involving group theory, representation theory, and category theory, can be used to formalize molecular structures in simplified form according to symmetries and structural equivalences. Such techniques are ‘simple’ both inasmuch as they are inexpensive and insofar as that they yield representations that compress potentially complex molecular components and systems into wieldy, formal expressions.

We can cite a precedent case for such work. Merkle (1993) introduces a mathematical proof of the isomorphism of potential energy levels of molecular bearings from the first principles of bearing symmetry. We can expand and improve upon this particular approach by using group representation theory and proof assistants such as Lean or Coq. Mathematical techniques may prove to be useful for designers who seek to decompose their designs into subcomponents (Hogg 1999, 305-306) or understand how such subsystems might compose larger systems. (Drexler 2006, 10) Zimmer (1990) describes such a technique (the application of category theory and group theory to decompositional and compositional models) as “representation engineering”.

Can existing silica infrastructure be used to interplay and boot-strap the first diamondoid structures?

I have seen many reports in which silicon and germanium play important roles. In certain instances, germanium has been identified as possessing particular qualities that render it uniquely suitable for certain applications. For instance, germanium has been nominated as the optimal element for the design of a hydrogen donation tool, given the (preferable) weakness of the bonds that Ge forms with H. (Freitas 2011, 393) To take another example, Freitas et al. (2007) employ germanium atoms for holding C2 dimers in tooltips. Therefore, at this stage, it is not clear (at least to me) whether or not silicon will prove to be used exclusively without other heavy carbon group elements.

With respect to bootstrapping diamondoid structures, I keep in mind that the number of possible theoretical diamondoid structures that can occupy a given cubic volume is astronomical. It is sometimes cited that 10148 different diamondoid structures per cubic nanometer are possible. (Drexler 1995, 326) However, this estimate is produced with adherence to some conservatism, which, if suspended, yields an estimate of 10217 per cubic nanometer. (Drexler 1992, 264) This second estimate is obtained by first calculating the number of configuration space options for a diamondoid structure from the entropy of fusion for silicon and germanium, which are roughly equal. (ibid, 263-264) Thus, the vastness of the theoretical combinatoria for Si and Ge diamondoid do not differ. Given such an astronomical number of combinations, “the probability that modern technology can fabricate a particular randomly picked structure from the set defined above is effectively zero.” (Drexler 1995, 326)

Thus, when considering the prospect of diamondoid bootstrapping, combinatorial explosivity will become manifest as we entertain different design choices. For instance, “[a] positionally controlled dimer could be bonded at many different sites on a growing diamondoid workpiece, in principle allowing the construction of a wide variety of useful nanopart shapes.” (Freitas 2011, 392-393) Therefore — if you will forgive my conservatism on this point — I will opine that further exploratory engineering and experimentation in diamondoid mechanosynthesis (DMS) must be conducted before the question of silicon can be settled.