GoKawiilThe Promise and the Gap In December 1959, Richard Feynman stood before the American Physical Society at Caltech and told a room of physicists there was "plenty of room at the bottom." He was inviting them to join a new mission of exploration: the deliberate control of matter at the atomic scale. Nearly seventy years of world-changing progress followed downstream of that invitation: modern electronics, Moore's Law, and our deepest grasp of physics itself. We learned to engineer the world atom-by-atom, and nearly everything we now call technology rests on that work. Richard Feynman American Physical Society Meeting "Plenty of Room at the Bottom" Caltech - December 29, 1959 Thanks to that work, the materials that will drive the next wave of technology — for AI, for quantum, for energy and electrification — are, for the most part, not waiting to be discovered. They are already known. They are valuable. They can be made in the lab — but they are stuck there. We cannot make them at production scale. Put simply, materials innovation has a scale-up problem, not a discovery problem. Every major technological shift begins in the physical world of materials. The intangible breakthroughs we celebrate — the model, the qubit, the grid — are all downstream from someone learning to manufacture a substance reliably, at yield, inside a real device. When that manufacturing stalls, the future stalls with it. “The breakthrough was not the material.
It was learning how to process the material at scale.” Intel provides a clear illustration. By the 2000s, the silicon dioxide that had insulated the transistor gate for four decades had been thinned to a few atoms across, and it was leaking. Intel knew it needed a high-k dielectric years before it could ship one. The hafnium-based material it landed on was not a eureka discovery and required synchronized changes to other materials in the stack to be effective. More than a decade of work went into making that material manufacturable: integrating it into a real transistor stack, at yield, without breaking everything around it, by depositing one atomic layer at a time. Intel "Penryn"-family Processors 45nm Hafnium-based High-k Metal Gate Wafer January 27, 2007 When Intel finally shipped it at the 45-nanometer node in 2007, Gordon Moore called it the biggest change in transistor technology since the late 1960s. The breakthrough was not the material. It was learning how to process the material at scale. The materials scale-up bottleneck is a bottleneck on our future.
Why This Problem Remains The hurdles to solving this problem are twofold: physical and informational. The physical difficulty is that materials do not exist in a vacuum. They live in context, nested inside heterogeneous device structures, one material grown onto another, each environment changing both what optimum looks like and the path to reach it. Nearly every degree of freedom in the design of material and process is continuous and interconnected. Simulation and digital twins can provide guidance but not reach the last-mile fidelity of a real material. So synthesis remains trial and error, guided by hard-won expertise and informed intuition. Tuning a synthesis process is like operating tweezers while wearing oven mitts. “That's the heart of it: the throughput of data generation has exploded, but our ability to use it has not.” The informational difficulty compounds the challenge. Characterization is fragmented across complementary, narrow probes — XRD, XPS, RHEED, TEM, AFM, and many more — each with its own hardware, its own software, and its own sub-specialty to separate the signal from the noise. Practitioners stitch the picture together by hand: serial, operator-biased, lossy. Metadata is left behind. Null and negative results, which carry real information, are routinely thrown away. Skilled engineers may feel that their analysis is already sufficient and the data captured is excessive, but nothing in the toolchain ever surfaces what's being discarded. Molecular Beam Epitaxy (MBE) Tool National Renewable Energy Laboratory (NREL) That's the heart of it: the throughput of data generation has exploded, but our ability to use it has not.
Why Now So why is this problem solvable now, when it was not before? Two forces have converged. Modern tools have become sensor-rich and high-throughput, producing real-time data in volumes that were unimaginable a decade ago. AI is now capable of using such data, having passed significant thresholds in compute, memory, bandwidth, and transfer learning that works on limited datasets rather than demanding oceans of examples. “We need systems that use the immense data that already exists to intelligently guide growths — instead of only explaining failures.” The honest remaining constraint is the same thing that makes the problem both hard and defensible: there is no "internet of materials" to scrape. The data that does exist is rich, but it is proprietary and siloed. A generic, web-trained AI model cannot solve this, and the opportunity isn't in waiting for more data — it's in building a system that finally puts to use the immense amount of data already being generated.
What Atomscale Does The bottleneck has never been a shortage of promising candidate materials. It is the decades of trial and error it takes to manufacture even one of them reliably. Atomscale turns that grind into a guided science. It reads the data your tools already produce and uses physics to turn raw signal into insight in real time — so you can steer a run while it is still growing, rather than explaining it after it has failed. Every run becomes legible. Every insight becomes cumulative. Knowledge compounds instead of dying in a silo. Today’s tools: siloed, single-use models Raw process and metrology data feeds isolated statistical, physics, and generic models whose insights dead-end with no carry-over across runs. Today’s tools SILOED, SINGLE-USE MODELS DATA MODELS OUTPUTS Statistical model Physics model Generic model Chat assistant Raw process & metrology data Most signals are lost Insights dead-end; no carry- over across runs Atomscale: a hierarchy of fit-for-purpose models Physics-formed adapter, time-series, and reasoning layers extract and record 43x more useful information and steer every run as it grows with real-time insights. Atomscale A HIERARCHY OF FIT-FOR-PURPOSE MODELS Extract & record 43x more useful information Steer every run as it grows with real-time insights Reasoning layer Material insights and action Time-series layer Capture signals over time Adapter layer Physics-formed extraction Atomscale's architecture uses a hierarchy of models. Physics-informed adapters sharpen signal-to-noise at the base and pass only meaningful information upward to time-series and reasoning layers. That hierarchy acts as a sandbox, constraining general-purpose AI to domain-relevant data, so the system behaves like a knowledgeable colleague instead of a generic LLM. Domain expertise is built into its structure, combining the leverage of automated reasoning with deterministic outputs grounded in physics. This is not a promise; it is already measurable. In encoding information about an unseen run, Atomscale performs 43x better than baseline unsupervised machine learning. Atomscale changes the daily texture of work. Before: manual analysis, fragmented tools, and insight trapped in one person's head. After: automated feature fingerprints, queryable data, runs steered in real time, recipes once too complex to maintain now within reach. This lifts the burden of operating today's processes while improving tomorrow's because knowledge no longer dies at the end of an investigation. It accumulates across the organization, on a level playing field. Engineers stop being the workhorses of their processes and instead become the managers. The whole organization can begin to aim for 10x gains instead of incremental ones. “We clear away the data drudgery so people do more of the thinking, not less. We augment the expertise that already exists.” This is not a replacement for engineers; on the contrary, there has never been a more exciting time to wield the levers of physics and develop exceptional real-world technologies. Design, conceptualization, creativity, the setting of goals and requirements — these stay human. We clear away the data drudgery so people do more of the thinking, not less. We augment the expertise that already exists. Feynman collaborating with prominent physicists Shelter Island Conference on Quantum Mechanics June 1947
What We Believe Our claim: the right AI today can extract far more meaningful information in aggregate from the data already collected than human analysis. Parts of the field may disagree. The teams that embrace it are the ones that will lead the coming wave of materials innovation. We're headed to a place where materials measurement data becomes a true asset: a physics-rich, interoperable space that reasoning agents can navigate and monitor continuously. Scale-up timelines will compress from a decade toward something far shorter. The materials trapped in the lab today will begin to reach the world on a human timescale. We drilled to the bottom decades ago. Feynman was right: there was plenty of room, and we have spent nearly seventy years proving it. The first phase of exploration is complete, and the next phase is different. We must learn to navigate what we found and bring it back up, from atom to scale. We invite you to join us on this path.