An Investigation into Atomically Precise Nanofabrication for a 3D Diamondoid Spatial Tensor Engine

1.1 Problem statement and research scope

The materials that make an era

The futures we imagine must eventually be made from something. A spacecraft needs a structure that can survive its journey. A fusion reactor needs materials that can confine, withstand, and extract energy from an extreme environment. An intelligent machine needs a physical substrate on which to compute. Behind each vision lies a quieter question: what must we learn to make before that future becomes possible?

Human history repeatedly turns on answers to this question. Shaped stone extended the capabilities of the human hand. Copper smelting made workable metal available; bronze alloying and casting enabled more durable tools and new forms of production. Iron and, later, industrial steel expanded the scale of machines, transport, and construction. Synthetic polymers made lightweight, moldable, insulating materials available in enormous quantities. Semiconductor purification, controlled doping, and precise fabrication helped turn silicon into the foundation of modern computing. 1

These transitions paired a material with a way to manufacture it. Knowing that bronze could be useful was different from controlling an alloy and casting it repeatedly. Silicon became transformative through an industrial system capable of growing crystals, patterning devices, and producing reliable circuits. A discovery changes the technological landscape when its properties can be delivered in a useful form, at a useful scale, and at an accessible cost.

Scientific ideas, algorithms, and institutions also drive progress. But every technology that acts in the physical world ultimately encounters the capabilities of its materials and manufacturing processes. This research starts from that recurring constraint: expanding what we can reliably build may expand what civilization can do. The futures that inspire an engineering generation become reachable through the tools that generation learns to create.

Intelligence and the physical demands of the future

Artificial intelligence brings this relationship into unusually sharp focus. Its apparent abundance in software depends on scarce physical infrastructure: accelerators, memory, advanced packaging, power systems, and facilities that can remove heat. In its 2025 assessment, the International Energy Agency estimated that data centres consumed approximately 415 terawatt-hours of electricity in 2024 and projected substantial growth under its baseline scenario. These are estimates for all data centres, rather than AI alone, but they show the scale of the substrate supporting the digital economy. 2

The ambitions described as artificial general intelligence and artificial superintelligence raise the stakes further. AGI would imply broadly capable systems; ASI would imply capabilities substantially beyond those of humans. Their feasibility and arrival dates remain unsettled. Yet a future in which increasingly capable intelligence is widely available would still require affordable computation. The consequential goal is access: enabling a student, a small laboratory, a physician, or a community to use capabilities that otherwise remain concentrated behind expensive infrastructure.

Efficiency can make that demand larger. Jevons’s paradox describes how reducing the resource cost of an activity can encourage enough additional use to outweigh the savings. Applied to computation, this is a possibility rather than an automatic law: cheaper intelligence could invite more users, more ambitious tasks, and entirely new applications. Lower energy per useful result therefore matters both for serving existing demand and for making a much larger range of activities affordable. 3

The same material questions run through our ambitions beyond AI. Sustained activity in space requires structures, shielding, thermal systems, and manufacturing suited to difficult environments. Better solar cells require control over light absorption, charge transport, interfaces, and long-term stability. High-field magnetic fusion depends in part on manufacturing reliable superconducting conductors; practical reactors also face demanding heat and radiation environments. Batteries require controlled electrode and electrolyte structures that balance capacity, transport, safety, and lifetime. These advances involve different chemistries, but each turns a desired function into a manufacturing problem. 4, 5

Medicine adds another frontier. Photonic devices control light; phononic devices control mechanical vibrations and their propagation. Alongside chemical and quantum sensors, such structures could expand our ability to detect small forces, local fields, and biological signals. Diamond-based nanosensing already provides experimental examples of coupling atomic-scale features to measurable signals. Future devices will depend on producing their active structures, interfaces, and surrounding systems with sufficient consistency. 6

AI can accelerate the search for those structures. GNoME demonstrated large-scale computational prediction of candidate crystals, while A-Lab demonstrated an automated synthesis workflow for selected inorganic materials. These are distinct achievements: predicting a stable structure, producing the intended sample, and verifying a useful property are separate steps. As design tools improve, the gap between structures we can propose and structures we can reliably manufacture could become increasingly consequential. Closing that gap is the broader motivation for this work. 7, 8

A bet on carbon

My bet is that carbon can provide a powerful starting point for closing it. The same element forms the layered structure of graphite, the sheets of graphene, the tubular structures of carbon nanotubes, and the three-dimensional covalent network of diamond. Changing the arrangement and bonding changes the available behaviour. A possible carbon age would be defined by learning to manufacture selected arrangements deliberately, so their capabilities become dependable engineering resources. I want to help build the infrastructure that makes this possible: a foundation on which future researchers can construct machines and materials that today exist only as carefully specified designs.

Graphene makes both the opportunity and the difficulty tangible. In 2004, Andre Geim, Konstantin Novoselov, and their collaborators isolated and studied atomically thin carbon sheets using mechanical exfoliation, famously involving adhesive tape. Their experiments opened a new field around a material with remarkable electronic and mechanical behaviour. Graphene became a striking demonstration of how an ordinary source material, graphite, could contain a radically different technological possibility when its structure was accessed at the right scale. 9

Translating that possibility into large, integrated products proved more complicated than reproducing a headline property. Growing a larger sheet can introduce grain boundaries; transfer and processing can add contamination, wrinkles, damage, and problematic contacts. The consequences depend on which property and device are being measured. Research on polycrystalline graphene, for example, found that grain boundaries could weaken membranes while affecting electrical behaviour less severely. Exceptional results from a carefully characterized sample do not automatically become the properties of every film, composite, or finished component. 10

This history makes precision an application-specific requirement. A coating or composite can deliver value with a distribution of flake sizes and defects. A device whose function depends on a particular atomic edge, junction, or moving clearance may require much tighter control. The manufacturing challenge grows as performance depends less on an average material property and more on the exact arrangement of the working structure.

Diamond offers a compelling structural starting point for the latter problem. Its carbon atoms form a stiff three-dimensional network of covalent bonds. High-quality diamond also combines valuable thermal, optical, and electronic properties, and fabricated diamond resonators have demonstrated very low mechanical loss under specified conditions. These features motivate its use in small mechanical structures; they do not establish the performance of every proposed molecular joint. 11

Diamondoids are small, rigid, diamond-like molecular frameworks. In the proposed manufacturing system, predominantly carbon and hydrogen structures would become links, supports, constraints, and other designed components. Their shape, bonding, and surface termination would determine how they interact. The attraction is a relatively restricted structural palette from which many geometries might be assembled. Other elements would still be needed for some tools, interfaces, and applications; carbon is the starting platform of this research.

From precise chemistry to a nanofactory

K. Eric Drexler developed the idea of molecular manufacturing into an engineering programme, including the analysis presented in Nanosystems in 1992. Robert A. Freitas Jr. and Ralph C. Merkle subsequently developed reaction proposals, assembly concepts, and research plans directed toward diamondoid manufacturing. Chris Phoenix also described a nanofactory architecture in which small fabrication units feed successive stages of assembly. Together, these proposals make a demanding claim worth investigating: a factory could organize molecular-scale operations into a process that produces useful objects with controlled atomic structure. 12–14

The basic idea is understandable without imagining a machine that manipulates arbitrary matter. Prepared molecular feedstock supplies suitable chemical groups. A positioned tool brings a reactive group to a selected site on a workpiece, under conditions intended to form or remove a particular bond. A sequence of such operations produces a component. Assembly stages then combine components into larger structures. The allowed products depend on the supported reactions, available geometries, and access to the growing workpiece. 13, 14

Scale would require many fabrication units operating in parallel. Small components would move through progressively larger assembly stages, much as parts move through a conventional factory. The design must also supply feedstock, prepare and renew tools, deliver power and instructions, remove waste and heat, inspect products, and recover from defects. One successful reaction is the beginning of this manufacturing loop. A productive nanofactory would need the loop to work repeatedly, with a yield and throughput that justify its cost. 13, 14

The experimental history contains genuine advances toward individual capabilities. IBM’s Eigler and Schweizer demonstrated positioning individual xenon atoms on a nickel surface at low temperature, reported in 1990. Rothemund’s DNA origami work in 2006 demonstrated programmable molecular self-assembly into designed nanoscale shapes. These achievements established different kinds of control: moving selected atoms on a surface and encoding assembly through molecular recognition. Neither, by itself, supplied a route to arbitrary three-dimensional diamondoid machinery. 15, 16

Industrial efforts also explored the missing bridge to scale. Zyvex proposed exponential assembly using many miniature robotic arms, and DARPA’s Atoms to Product programme targeted assembly across nano-, micro-, and larger scales. These programmes document attempts to connect small structures to useful products. The nanofactory architecture examined here remains a proposal rather than an established production platform. Its history should be read through demonstrated capabilities and unmet objectives, rather than treating every nanoscale product as evidence that the whole system works. 17, 18

More recent experiments bring positionally controlled carbon chemistry closer to the centre of this programme. A 2026 preprint by Cowie and colleagues reports carbon-pair donation to prepared sites on hydrogenated silicon, patterned placement, and stepwise construction of carbon chains using inverted-mode scanning tunnelling microscopy. This is evidence for specific chemical operations on a prepared surface. Building a rigid diamondoid framework, assembling internal moving parts, and producing many such devices remain further steps. 19

There are also instructive failures within candidate computing designs. Seet, Ouldridge, and Doye found that pipelining the molecular mechanical circuits they studied could increase both errors and irreversibility. A mechanism that behaves well in isolation can lose that behaviour when driven alongside connected components. Such results identify where the proposal must become more precise: reaction selectivity, access, synchronization, lifetime, inspection, and cumulative yield need experimental or validated computational answers. 20

Why a spatial tensor engine

That ambition needs a concrete application capable of exposing its hardest requirements. The proposed 3D Diamondoid Spatial Tensor Engine serves that purpose: an arithmetic architecture in which controlled molecular motion represents information, and connected mechanisms transform and transmit it through a three-dimensional structure. In a spatial architecture, operations occupy physical cells and pass information between neighbours. A tensor engine would arrange those cells to perform the repeated multiplication and addition underlying matrix and tensor workloads.

Graphene synthesis and other carbon-material applications were considered as starting points. Many can become commercially useful through improved growth, purification, formulation, or integration, with some defects tolerated. Chemistry can already produce many atomically precise molecules, and self-assembly offers additional routes to ordered structures. The engine selects a more demanding target: specified relationships between constrained moving parts, where geometry determines whether arithmetic works at all.

The engine is selected because it makes atomic precision a functional requirement at critical sites. Its clearances, supports, couplings, and signal paths must establish specific relationships between constrained moving parts. Correctly synthesized components must then preserve those relationships when assembled into a circuit and operated repeatedly. This tests whether precise fabrication can deliver a working machine as well as its individual parts. Manufacturing many copies would additionally require coordinated production and assembly. A nanofactory is one proposed architecture for delivering these capabilities; any successful alternative would have to meet the same structural and production requirements.

Imagine the distant endpoint: a diamondoid engine the size of a sugar cube, cool enough to hold, performing computation on a scale we now associate with the world’s AI data centres. That is the arresting possibility motivating this investigation. If manufacturing and low-loss operation could support such an extreme concentration of useful computation, access to intelligence could change as profoundly as access to computing changed with the integrated circuit.

The theoretical bridge to that possibility rests on molecular dimensions, energy recovery, and the dissipation of particular mechanisms. Small components could permit a high density of working elements if space for their supports and interfaces can also be accommodated. Information-preserving logic, paired with an appropriately controlled physical cycle, could reduce the energy discarded during computation. Hogg and colleagues used molecular dynamics at 300 K to estimate low rotational drag for a hypothetical joint built from acetylenic bonds within a rigid diamondoid housing. Together, these ideas provide a concrete reason to investigate dense, low-loss computing. The joint result is a simulation of a specified structure; extending it to a driven processor is the next scientific question. 21, 24

The image is a vision, not a demonstrated device or an established comparison with the combined performance of today’s accelerators. Its validity would depend on arithmetic precision, clock rate, reliability, memory, communication, drive recovery, and heat removal. Earlier molecular mechanical computing proposals motivate investigating very high density and low dissipation, but a processor must establish those advantages at a complete-system boundary. The research begins with a small, explicit arithmetic operation and asks whether its promise survives each step toward a working machine. 21

A working engine could be consequential twice: through the computation it delivers, and through the manufacturing capabilities required to create it. The same investment in reaction control, assembly, and verification could establish a starting point for other molecular machines. That is the larger invitation behind this thesis—to give researchers, engineers, and backers a concrete path into a frontier whose successful development could shape what subsequent generations are able to build.

The energy required to compute

The energy required to compute, move data, and remove heat constrains the scaling of modern computing systems. Increasing device density yields useful performance only when power delivery, communication, and cooling can support the resulting workload. This is where the broader manufacturing ambition becomes a specific computing problem.

Modern digital processors predominantly use complementary metal-oxide-semiconductor technology, or CMOS. A logic state is represented by a voltage on a circuit node, with charge stored in its effective capacitance. Transistors act as voltage-controlled switches that charge or discharge these nodes. In conventional switching, energy is dissipated through resistive paths; leakage and other losses add to the operating cost. Data movement through wires, memory, and interfaces can contribute substantially to the energy required for useful computation.

Landauer’s principle places a minimum thermodynamic cost on logically irreversible information erasure under its stated conditions. It does not require every computational transition to dissipate that amount. Logical reversibility preserves information; appropriately designed physical processes can seek to recover energy. Reversible and adiabatic electronic approaches also pursue this goal, so the relevant test for molecular mechanics is the performance of an alternative implementation of useful computation. 22, 23

In the proposed engine, distinguishable molecular configurations would encode information, while mechanical couplings would coordinate state changes. A carefully controlled drive could, in principle, return part of the energy used during a transition rather than discard it as heat. The challenge is to maintain that behaviour at a useful rate, in a thermally fluctuating environment, across many connected components. Stiffness, small size, and reversible logic alone cannot settle those questions. 20, 21

This research therefore connects two interdependent domains. Nanomechanical computing architecture defines the representation of information, arithmetic, timing, and system interfaces. Positionally controlled molecular manufacturing defines the reactions, tools, assembly operations, and verification needed to construct that architecture. Device design establishes fabrication requirements; fabrication feasibility constrains which designs can become physical machines.

The broader ambition is therefore approached through a defined computing problem. A useful result must establish both that the molecular architecture performs reliable arithmetic and that a reproducible process can construct the structures it needs. The central research question makes those two requirements explicit.

1.2 Central research question

Can rigid molecular mechanisms be assembled into a reliable, clocked arithmetic system whose energy efficiency, spatial density, and error rates justify a dedicated atomically precise manufacturing process?

The question is evaluated along two coupled tracks: device utility and manufacturing feasibility. Device utility requires an advantage at a defined workload, precision, throughput, and error rate. Manufacturing feasibility requires a reproducible route to the structures that deliver that advantage. A positive result on one track is insufficient to establish the complete system.

1.3 Scope of success

The first benchmarks are an input-preserving half-adder and a two-term dot product with two-bit unsigned inputs, specified in Chapter 3. They provide bounded tests of arithmetic, retained information, timing, and composition before a larger tensor architecture is evaluated. The decision gates in Chapter 11 connect these specifications to physical modelling, driven circuits, fabrication, experimental verification, and system scaling.

Evaluation proceeds from a component to a circuit, an architecture, and ultimately a fabrication platform. At the component level, the task is to characterize intended motion, unwanted displacement, and dissipation under specified operating conditions, initially including a room-temperature reference of 300 K. At the circuit level, the task is deterministic arithmetic with an explicit drive, synchronization, and readout protocol. At the architecture level, the task is useful numerical work with storage, communication, and interfaces included. At the platform level, the task is to determine whether the fabrication operations extend to additional structures within a defined chemical and geometric scope.

These levels establish the claims the investigation must support. The present manuscript is an end-to-end hypothesis thesis; performance estimates, simulations, and experimental measurements will be recorded with their methods and uncertainty as the programme develops.