Material Progress: Developing AI's Scientific Intuition, with Orbital Materials' Jonathan & Tim
Summary
Orbital Materials’ core bet is that AI can turn materials discovery from artisanal trial-and-error into a more predictable engineering discipline. Its generative models propose candidates, force-prediction models qualify them, and an internal wet lab closes the loop. Jonathan Godwin says that by the time they decide what to make, they have answered roughly 90% of the questions needed for confidence and narrowed the field to two or three materials.
The deepest technical signal is out-of-distribution generalization from roughly 20-atom inorganic crystals to a complex protein system. Jonathan calls it evidence that the models are “learning something really fundamental at that small scale.” Message-passing neural networks do not use explicit positional embeddings like language models, allowing atomistic simulations to grow with available computing resources, subject to practical limits.
The potassium-channel work offers a striking but explicitly unconfirmed demonstration of that generalization. On one V100 GPU, Tim Duignan observed water entering the channel and a hydroxyl group drawing it in through a hydrogen bond—a candidate explanation for mutation experiments in which removing that group reduces conductivity by almost an order of magnitude. The work had not yet been peer-reviewed or experimentally confirmed; Tim describes it as a strong, plausible hypothesis.
Orbital’s initial commercialization focus is the data-center buildout rather than moonshot materials such as room-temperature superconductors. It is developing thermal-management and decarbonization materials, including a CO2-capture material designed to use data-center waste heat and airflow. The economic fulcrum is regeneration: one-off capture is relatively solved, while lowering desorption energy and reusing a material for perhaps 10 years could radically reduce costs.
The model economics resemble image diffusion more than frontier language models. Orbital says it has built leading models and a wet lab with less than $40 million raised. Atomistic simulations can scale across additional chips because the architecture has no Transformer-style positional limit, though large systems become slow, long-range electrostatics may require additional modeling, and density functional theory is inadequate for some highly correlated electron systems.
The defensibility may come from combining scalable models with tacit experimental knowledge that has historically been difficult to copy. Jonathan compares semiconductor process expertise—perhaps 10,000 steps, transmitted through a master-and-apprentice process—with TSMC’s accumulated advantage. Orbital’s internal lab therefore matters less as manufacturing capacity than as a fast feedback system that keeps its AI grounded in manufacturable materials.
Higher scientific output may come with a painful redistribution of satisfying work from humans to machines. A study summarized by the host reported 44% more materials discovered, 39% more patent filings and 177% more downstream product innovation, while 82% of scientists reported lower job satisfaction as AI handled more idea generation and humans did more validation. Jonathan’s forecast is that “creativity is one of the first things that’s going to go,” even as he argues accelerated discovery is necessary to produce more for less.
Deep dive
1. Materials set the physical ceiling on the AI economy
Jonathan’s semiconductor example starts beneath the software stack: microscopic debris can disrupt etched circuitry, so insulating low-k materials must fit between increasingly narrow features and protect the chip. Without significant advances in those materials, “you don’t get more transistors on your chip and you don’t get AI.”
His broader framing is deliberately maximalist: “Materials science is the original tech startup.” The Silicon Valley revolution rests on materials, and the next frontier—from better chips and GPUs to spacecraft and new optical materials—makes materials science “the fundamental missing piece in achieving the science-fiction future.”
Tim’s inverse test is to name a major problem and trace it back to materials. Climate mitigation is the clearest specimen: batteries and solar panels produced massive CO2 reductions through materials breakthroughs, while most tools for addressing climate change are, in his broad definition, materials-science problems too.
2. Discovery still depends on luck, overdesign and accumulated intuition
Jonathan compares present-day materials development with bridge building before reliable structural modeling. Engineers could build remarkable bridges, but uncertainty forced overdesign—Brooklyn Bridge is his example—and occasionally produced catastrophic failures; modern software instead resolves the forces, stresses and limits on every bolt and beam before construction.
The desired transition is equally concrete: scientists should know that a proposed material will have the required properties before making it, then use computation to diagnose disappointing yield or performance “from the ground up,” including what is happening at molecular scales that experiments cannot directly reveal.
Jonathan’s semiconductor example shows why this is hard. A chip may involve something like 10,000 process steps, with specialists spending 30 years learning details such as which gases clean and etch a wafer; the knowledge is often unwritten, not reducible to basic chemistry and passed down in a master-and-apprentice fashion.
That tacit corpus helps explain why money alone cannot instantly recreate TSMC. Tim nevertheless resists treating AI as simple replacement: human and machine intuition arise differently and may catch one another’s mistakes, even if moving beyond the human brain’s limits means computationalizing some “beautiful” artisanal knowledge.
3. Orbital combines models and a wet lab because neither side closes the loop alone
Jonathan founded Orbital after leading large-scale materials-AI work at DeepMind. Capabilities he expected to take a decade arrived much faster: calculations impossible on the best supercomputers 10 years earlier were becoming laptop tasks, convincing him that materials AI was following the broader field’s steep capability curve.
The organizational thesis was as important as the model thesis. An AI lab would not naturally assemble an experimental materials team, while traditional chemistry and materials companies had spent “the last 25 years” saying no to computational tools; a new company could unite frontier modeling with scientists responsible for making and commercializing actual materials.
Jonathan’s endpoint is a materials-design interface with an LLM copilot: request a battery or other functional material, iterate computationally, and reach high confidence that the laboratory product will possess the predicted properties. The wet lab grounds that ambition in synthesis constraints and supplies rapid feedback when simulations fail.
4. Neural networks appear to compress physics into usable intuition
The host’s foundational analogy is catching a ball: humans do not explicitly calculate every trajectory, yet behave as if they possess an intuitive physics. He sees similar shortcuts appearing across protein folding, atomic simulation, weather and other domains, and asks whether networks are learning compact higher-order concepts rather than merely accelerating equations.
Tim agrees while admitting, “I don’t think we fully understand why yet.” His candidate explanation is dimensionality reduction: physical models succeed by retaining consequential variables and discarding irrelevant ones—electrons in one simulation, solvent in another, or the microscopic air dynamics around a ball.
This connects to coarse graining and statistical mechanics: behavior can repeat across scales, and accurate prediction need not track every component. Tim’s formulation is the episode’s through-line: “Finding efficient ways of keeping track of the important information and losing the unimportant information” is a central problem that machine learning handles unusually well.
Jonathan points to architectural universality: similar networks support materials simulation and weather forecasting, while compact world models reproduce water, reflections, gravity and small visual effects formerly requiring expensive rendering. They make “shortcuts that preserve the fidelity,” like extraordinary athletes performing feats without knowing the mathematics underneath them.
5. First-principles simulation turns data scarcity into a scaling opportunity
Tim was initially skeptical that AI could help electrolyte chemistry because experimental databases contained vanishingly little information—sometimes only a few parameters per electrolyte. The update came when he recognized that first-principles computation could automatically generate millions of data points, with active-learning loops concentrating new simulations where the model was weakest.
Jonathan distinguishes useful inference from feasible training. A realistic CO2-to-fuel catalytic process might require thousands or millions of atoms across long time spans, but training can use many roughly 100-atom fragments to learn local rules, then apply those rules to the otherwise intractable large system.
The result that “completely blew my mind” was training on inorganic crystals of roughly 20 atoms and then simulating a protein out of the box. Jonathan reads that as evidence of fundamental small-scale learning; Tim’s more tentative interpretation is that molecular interactions may contain only a finite amount of relevant complexity.
6. The foundational model and Orb make generation and dynamics two views of the same physics
The transcript refers to Orbital’s foundational model as “Linnaeus” in the host’s setup and “Linus” in Jonathan’s later explanation. Jonathan describes that foundational model as learning simplified physics through diffusion, with Orb models fine-tuned for physical simulation.
Tim traces diffusion models back to nonequilibrium statistical mechanics and Langevin dynamics. In a materials generator, atoms travel along a learned score vector field; in molecular simulation, they move under a force field. He cites work showing that when diffusion is trained on a simulated trajectory, the score can be mathematically identical to that force field.
Because a structure can be corrupted with effectively limitless noise samples, diffusion training provides dense coverage of simplified physics. Fine-tuning against forces computed through quantum calculations then produces models that can be used for physical simulation.
Noising a crystal largely means jiggling atom positions until the arrangement becomes incoherent. The unusual wrinkle is periodicity: an atom leaving one face of the simulation box reappears at the opposite face, requiring mathematical adjustments but leaving the overall procedure closer to image diffusion than protein-specific noising.
Atom identity is softened from a categorical choice among species 0 through 117 into a continuous prediction and then cast back. Orbital’s message-passing neural networks are trained on small crystal structures and do not use explicit positional embeddings like language models, allowing the simulated system to extend with available compute rather than hitting a fixed positional context length.
7. Generation proposes structures; force prediction makes them scientifically useful
A scientist could ask the foundational model for 10 materials with a specified band gap, much as an image model receives a prompt. Jonathan prefers exact numerical conditioning to natural language because materials often require a narrow “Goldilocks” interval: not merely more or less of a property, but a specific quantitative target.
Orb receives atomic coordinates and species, then outputs forces and energies. Decades of molecular-simulation methods can turn those primitives into diffusivity, chemical equilibria, binding strength and other foundational properties, making the apparently narrow force-prediction task a gateway to a broad materials toolkit.
The two systems form an outer loop: a generator creates hypotheses, the simulator subjects them to more exact tests, and the results help determine which candidates should reach the lab. Jonathan places the resource profile near CNN-based image diffusion, not an LLM; Orbital built models and a lab with less than $40 million raised, while he suggests the field may be at a “GPT-1” stage with substantial scaling headroom.
8. Simulation becomes a microscope, an experiment and a physical-data foundry
Tim’s first use case is a “computational microscope” for nanometers and nanoseconds—the billionth-of-a-meter and billionth-of-a-second regime where bonds form and break. As with the first microscopes, sufficiently broad simulation might yield discovery after discovery before researchers know what hypothesis to pose.
Dynamics are particularly valuable after static prediction. Protein structures may be available, but scientists still need to see how proteins move and interact below experimentally accessible microsecond scales; materials likewise require transition states, reaction pathways and evolving structures, not just final equilibrium snapshots.
The second use case is the in-silico experiment: change an atom type in an input file instead of ordering chemicals, synthesizing a compound and using microscopy to understand it. If simulations become sufficiently fast and reliable, thousands of candidate experiments can be explored before committing scarce laboratory time.
The third is a “bedrock for physical simulation.” Atomic runs can supply parameters for climate or other larger-scale models, generate training sets, and support a statistically and mechanically rigorous coarse-graining procedure—for example, ignoring solvent while learning effective free energies that reproduce the behavior of the components researchers care about.
9. Orbital’s workflow tries to eliminate 90% of uncertainty before synthesis
Jonathan illustrates the process with a metal-organic framework, envisioned as a sponge with large pores and many attachment sites. A generative model might propose 10, 15 or 20 candidates whose ligands act like tendrils, selectively grabbing desired molecules from a mixed stream.
Generation cannot conveniently encode every practical question. Simulation therefore checks adsorption kinetics, whether the tendrils obstruct pore space and other failure modes; Jonathan says, “By the time we get to making a decision about what to make, we’ve answered 90% of the questions” required for confidence.
Unexpected simulated behavior feeds another creative round. After two or three iterations, Orbital may take two or three computationally qualified structures into the lab—candidates with a much higher expected success rate and enough novelty to reach chemical space a human team would not ordinarily explore.
Jonathan treats novelty as nearly as important as reliability. The AI should not merely reproduce established chemical intuition more quickly; its value is to expose unfamiliar designs, while simulation and experimental work test whether those designs are physically and practically viable.
10. The potassium channel tests whether crystal-trained physics transfers to biology
Tim begins with the body as “really an electrical device.” Cells use ions rather than electrons to create voltage changes; potassium channels control signals underlying muscle contraction, heartbeats and neuronal firing by selectively allowing potassium to flow through cell membranes.
Yet the selectivity filter remains a keystone unsolved problem: researchers still dispute whether potassium passes through alone or accompanied by water. Experimental evidence is mixed, while classical simulations generally assume that potassium ions pass through alone, signaling that the field may lack a crucial observation or modeling capability.
After Orb performed well on electrolyte solutions, Tim asked whether it might be the universal force field computational chemists had wanted for decades. He expected an out-of-distribution protein simulation to “crash immediately” or become unphysical; instead it remained stable, exposed interesting behavior and left him “literally…jumping up and down.”
The training distribution contained small, highly repetitive crystals rather than membranes or proteins. Tim’s tentative explanation is that a finite set of information—such as atomic dipoles and spatially decaying interactions—may suffice; Jonathan attributes the result in part to scale and AI engineering applied to materials.
11. A hydroxyl group may resolve the water-conduction dispute
Tim’s caveat is load-bearing: the work had not been peer-reviewed. Potassium-channel specialists considered the result interesting, plausible and consistent with experiments, but direct experimental confirmation was still needed; the simulation should therefore be read as a precise, testable hypothesis rather than an established mechanism.
Orb showed water entering the channel’s narrow selectivity filter, unlike many classical simulations. A hydroxyl group on the protein reached toward the water, formed a hydrogen bond and appeared to pull it inside—a new qualitative phenomenon Tim had not seen reported elsewhere.
The crucial corroborating fact is mutation evidence: removing that hydroxyl group lowers channel conductivity by almost an order of magnitude. When the host asks whether the new mechanism likely explains that result, Tim says that if forced to bet, “I think it definitely is,” while preserving the need for validation.
The medical path runs from mechanism to mutation. Defective potassium channels contribute to heart conditions including long QT syndrome and heart palpitations; simulations might reveal how mutations impair current and eventually guide restorative drugs, an approach Tim notes has already been pursued for chloride channels.
12. Scaling is architecturally open but physically and computationally constrained
Tim’s initial run used one V100 GPU, so he cut the system off about 20 angstroms from the selectivity filter—roughly four or five atomic layers—and froze the outer layer. It was “a little bit of a hack” that stabilized the region long enough to inspect its central dynamics without simulating the full membrane.
A full system needs the channel, its lipid membrane and enough surroundings to avoid an infinite regress of stabilizing atoms. Periodic boundary conditions can close that world, but large systems become slow and eventually require many GPUs; reaching longer time scales is the main practical problem.
Jonathan says there is no language-model-style context ceiling because the network lacks positional encodings tied to a fixed training length. Predominantly local molecular interactions permit simulation size to rise roughly with added chips, although long-range electrostatic interactions decay as 1/r and may require longer-range models.
Coarse graining offers a route to whole-cell-scale dynamics by retaining selected components and replacing the rest with learned effective free energies. Neural-network potentials have produced very large simulations, but short accessible time spans and the difficulty of validating such systems remain central constraints.
13. Room-temperature superconductors expose the theory’s current boundary
Pressed for a room-temperature superconductor, Jonathan says, “Give us at least five years, but it may happen sooner,” while placing carbon removal and efficient AGI scaling first.
Tim explains the technical limitation: superconductivity is a highly correlated electron problem, not a predominantly local atomic one. Orbital’s current level of theory, trained on density functional theory, is not accurate enough to describe all the relevant quantum behavior, so the existing approach cannot fully solve the problem.
The tools could still contribute by proposing crystal structures or helping determine how to manufacture improved superconductors once researchers understand which electronic properties to induce. AI acceleration does not erase the fidelity limits of the physical theory used to produce training labels.
14. Data centers provide the first commercial proving ground
Orbital is prioritizing materials for denser, hotter AI infrastructure: rising rack and chip power densities demand better thermal management, while more GPUs require continued semiconductor improvement. Jonathan couples that opportunity to decarbonization because he views scaling AGI sustainably and cheaply as one integrated materials problem.
The product already discussed is a CO2-capture material designed specifically for data-center operating conditions. It can use waste heat and airflow to increase capture performance and potentially offset part of operators’ scope 2 and scope 3 emissions; Jonathan says more information about products will come over the course of the year.
Jonathan confirms the broad sponge-like mental model: reactive sites selectively interact with CO2 amid ordinary air, while adsorption capacity is tailored to operating temperature, air composition and other conditions. Regeneration is decisive—“the one-off capture is relatively solved”—because waste heat can lower desorption energy and a material lasting perhaps 10 years can spread its manufacturing cost.
Orbital does not expect to build “a massive plant somewhere in New Jersey.” Licensing and supplier relationships can separate product ownership from fabrication, as Jonathan compares with Nvidia’s model; meanwhile, its internal lab costs more than a software business but remains affordable for a venture-backed company and is indispensable for rapid experimental feedback.
15. AI shifts scientific labor toward validation—and could hide more reasoning
The host summarizes a study of AI-assisted materials research reporting 44% more materials discoveries, 39% more patent filings and 177% more downstream product innovation. The uncomfortable companion result was that 82% of scientists reported lower satisfaction as machines assumed more idea generation and humans inherited more validation work.
Tim experiences the tools as empowering because, as a physicist and mathematician without decades of chemistry training, simulations let him see bonds and mechanisms directly. He sympathizes with experts who spent careers building intuition but believes the satisfaction of getting impactful materials into the world more quickly can outweigh the loss of solving every problem alone.
Jonathan is less optimistic about preserving enjoyment. He says AI researchers already experience less satisfaction despite more prestigious, better-paid roles because their work has shifted from creativity toward engineering and expensive compute limits the experiments they can run individually. His forecast is that “creativity is one of the first things that’s going to go,” followed later by manual tasks and robotics.
Tim’s closing technical thesis is that LLMs plus neural force fields could pair hypothesis generation with an internal physical check. Models such as o3 illustrate the value of domains where answers can be validated, and neural-network potentials might let scientific systems test hypotheses computationally rather than sending everything to a laboratory.
Jonathan expects shared multimodal embeddings, though perhaps with separate decoders for different modalities. The host worries that reasoning in continuous latent space will be difficult to interpret and that AI systems may communicate more effectively with one another than with humans. The sobering fallback, quoted by Jonathan, is to “hope moral realism is true.”