🔬 Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk

The host opens with Periodic’s website claim, “intelligence is necessary but not sufficient,” and asks why that idea is less obvious than it sounds. Liam Fedus explains that no one can think their way to a solution in a universe as complicated as ours. Scientific progress requires forming conjectures and testing them through an iterative loop; rereading papers or thinking in isolation cannot anticipate every experimental outcome, including unexpected ones. Periodic therefore combines AI systems and simulations of the physical world with a physical, high-throughput laboratory, hoping that a different kind of intelligence emerges from their interaction. Fedus and Ekin Dogus Cubuk describe Periodic as an unusual combination of solid-state chemists and physicists, experimentalists and theorists, hardware engineers, LLM researchers, and computer scientists. Some enabling technologies, including robotic high-throughput experimentation and modern force fields, have only become practical in roughly the last three years. The closest historical analogy they offer is Bell Labs, where theorists, experimentalists, and chemists worked together. Their ambition is not merely to build a better digital model, but to create a lab whose data and experiments define the ultimate test of whether an AI-generated scientific idea is correct.

The host asks how working at Periodic differs from working in a large AI lab such as OpenAI or Google. Fedus says that Periodic’s reinforcement-learning environments are derived from its physical laboratories, and the lab’s measurements—not textbook or paper answers—are the final source of truth. Physical experiments introduce uncertainty at every stage: furnace labels can be noisy, machines can disagree, telemetry can be incomplete, and the actual temperature may differ from the intended temperature. In a mathematical or coding environment, the relevant axioms, functions, and APIs can in principle be made available to the model, and optimization has relatively precise answers. In a lab, an agent must make decisions from noisy, sometimes aberrant measurements while using a limited number of samples. The team therefore cares about decision-making under uncertainty, sample efficiency, variance reduction, and avoiding mismatches between training and inference. They cannot simply generate unlimited rollouts as a digital system can. Fedus describes the discovery loop as several linked tasks: determine whether a proposed atomic configuration is stable and has the desired property, find a synthesis process that can actually produce it, and characterize what was made. Since the result does not arrive labeled, each of these tasks needs its own AI environment and reward structure.

Cubuk explains that science requires dimensional reduction in a way that ordinary coding and mathematics do not. A physical system contains more atoms than any computer can represent, so researchers need reduced descriptors such as energy, energy fluctuations, reaction barriers, local atomic neighborhoods, temperature, and pressure. X-ray diffraction is another lossy projection: it does not reveal the complete structure, but gives important components or fingerprints from which a scientist must infer what is present. The experimental side has its own hidden variables. A furnace degrades because evaporated material coats its heating element, and its temperature is not perfectly uniform, so the location of a sample changes the result. Cubuk recalls a PhD laboratory where laser measurements varied depending on someone walking in an upstairs room and creating vibrations. He says this uncertainty, missing context, and noise are common in real-world intelligent work, yet much AI progress has focused on cleaner math, coding, and theoretical computer-science tasks. Fedus gives the example of X-ray diffraction characterization. The system must identify the phases actually present, reward chemically plausible explanations, and penalize spurious phases. It can also use experimental lineage and prior evidence to construct reinforcement-learning tasks that cannot be faked by a model that has merely memorized a published answer.

The host asks what a phase transition means and why it is useful as an AI signal. Cubuk uses familiar examples: ice remains solid as temperature rises until a threshold, when it abruptly becomes liquid; in the Ising model, spins remain apparently disordered at high temperature but align when the temperature falls below a critical point. In Periodic’s lab, researchers mix precursor crystals and change temperature or other conditions so atoms react into a new crystal. A major practical problem is that experiments rarely produce a clean yes-or-no result. A sample may contain unreacted precursors, amorphous material, and unexpected phases alongside the desired one. X-ray diffraction patterns can reveal these phases, but they do not always uniquely identify them. Different structures may be consistent with similar patterns, so the AI has to apply chemical and thermodynamic priors, compare results across replicates, and combine multiple characterization modes. Electrical, magnetic, morphological, and microscopy measurements can distinguish phases that look similar in XRD. Fedus describes this as an advantage for AI: it can combine more instruments, experiments, and temporal context than a human can comfortably track. The goal is often to “hill climb” phase purity. If the target phase appears at 1%, the system should recognize it, then choose conditions that increase its proportion and improve the evidence about what was actually made.

The host asks why materials science remains unfinished if most physical laws are already known. Cubuk says the famous idea that Dirac solved quantum mechanics and “the rest is chemistry” is misleading. High-temperature superconductivity remains unexplained, particularly unconventional superconductors such as cuprates that operate around 77 or 93 Kelvin and do not follow the simple isotope-effect physics of conventional superconductivity. Strong electron correlations also remain difficult for density functional theory to capture. The speakers invoke Phil Anderson’s phrase “more is different”: systems made of many components can display qualitatively new, emergent behavior that is not obvious from the laws governing a single atom. This is why perfect crystal models do not settle questions about real materials, whose microstructure, defects, interfaces, and collective behavior matter. Cubuk explains density functional theory as a way to replace an exponentially large quantum wavefunction problem with the more manageable charge density of the ground state. In principle the Hohenberg–Kohn result can be exact, but practical exchange-correlation functionals and kinetic-energy approximations introduce errors. Even a perfect DFT method would not eliminate the laboratory, because no computer can represent all roughly 10²³ atoms in a macroscopic sample. Simulations and experiments are therefore complementary: simulations scale quickly and estimate quantities such as formation enthalpy and stability, while experiments provide the real-world test and expose effects simulations omit.

Periodic uses simulations to filter and prioritize candidates, but the speakers stress that DFT cannot be trusted as the only source of truth. It generally models a perfect crystal, whereas real materials have microstructure, and it cannot readily predict every important property, including some superconducting behavior. Humans and AI use heuristics, experimental calibration, and physical evidence to decide which of many simulated candidates should actually be synthesized. The lab itself is being automated around data quality. Fedus says early experiments became bottlenecked by technicians and scientists manually operating equipment and inspecting morphology. Simple programs that captured a field of view and zoomed in were not intelligent enough, so Periodic began putting AI directly on instruments. These systems know the experiment’s objective, the intended material, and relevant prior evidence, allowing them to capture richer and more useful data at the moment of measurement. The team does not treat full autonomy as the goal. Its goals are large quantities of high-quality, diverse data, with automation used in service of those goals. Humans remain valuable for difficult dexterity tasks and for guiding campaigns, while routine bottlenecks are automated. One example shows the benefit: an AI system noticed a cyclic inconsistency across experiments, inferred that reversing a permutation made the data coherent, and led the team to discover that a machine had been loaded incorrectly. Standardized workflows should reduce the noise floor that often makes published experimental data less reliable than simulation accuracy.

The host asks how Periodic uses failures when successful materials are much rarer. A negative result can mean that the intended structure was not produced, that an impurity phase destroys the desired property, or that an unexpected structure emerged instead. Cubuk notes that the literature is biased toward positive results: researchers tend not to publish every failed synthesis, even though negative examples are essential for training a classifier. In-house experiments provide failures in the context of exactly what was attempted. The team also values a sequence of failures that eventually leads to success. That sequence records process engineering knowledge: which conditions were changed, how the evidence was interpreted, and how the researchers reached the positive result. The same known material may be difficult to reproduce, so the method and experimental lineage matter as much as the final label. Periodic aims to preserve conversations, intuitions, code, simulations, instrument outputs, and physical actions rather than training only on the final published result. Fedus says this produces reinforcement-learning environments based on historical states of the world: given the evidence available at a particular date, what choice did the scientist make and what happened next? This avoids a failure mode in which a pretrained model already knows the answer, fakes the reasoning, and learns strategies that do not generalize to novel materials. Even a future frontier model will need experiments because scientific discovery concerns what the model has not already been trained on.

Periodic plans to scale both laboratory capacity and computation, eventually building a network rather than relying on one Menlo Park site. Geography affects experiments, while local universities and cities provide different expertise. The company is also developing hardware rather than relying permanently on off-the-shelf instruments: a weighing step, robotic resting position, or contamination risk can become the bottleneck as throughput rises. The speakers describe this as a hardware-engineering problem as much as an AI problem. Its near-term commercial route is deployment into technically sophisticated industries, especially semiconductors. Periodic’s four-deployed engineers and researchers work on-site in secure environments, integrate AI and computational tools into customer workflows, and train systems locally on partner data. The company sees a progression from AI copilots to systems priced around outcomes, analogous to the way coding tools have moved from assistance toward more autonomous software production. It also contributes to open-source projects and supports academic groups through grants. For superconductivity, the speakers say the central challenge is not a shortage of chemical ideas but synthesizing and characterizing them. They cite Bednorz and Müller’s cuprate discovery, later work on nickelates, and magnesium diboride, which was found after a Japanese group screened roughly 30,000 materials and had sat unnoticed as a precursor for decades. Automated experimentation could dramatically increase the search surface: instead of trying 30,000 materials over a career, a lab might try them in a month. Periodic’s broader ambition is “synthesis superintelligence”—an AI-lab system that can continually propose, make, characterize, and learn from materials in the physical world.