Periodic Labs

Periodic Labs builds AI and laboratory systems for materials discovery. Its premise is that an AI model can propose a promising material, but an experiment must still make it and determine what formed. The company combines models, simulation and laboratory measurements to guide the next attempt.[1][2]

Liam Fedus and Ekin Doğuş Çubuk co-founded Periodic around that experimental loop. Fedus describes the AI challenge of learning from noisy and time-consuming physical feedback; Çubuk brings a focus on materials, synthesis and the limits of what an experiment can observe. Their stated ambition is to build systems that help researchers work toward new materials, including better superconductors.[1][2]

A search begins by proposing an arrangement of atoms that might be stable and have a desired property. The next problem is a recipe: which ingredients and processing conditions can actually produce it? After synthesis, a sample must be characterized. It can contain a mixture of phases rather than the single intended crystal.[1]

Periodic uses X-ray diffraction to help identify those phases: the spacing of atoms produces a pattern that acts as a structural fingerprint. Fedus describes early AI work that rewards correct phase identification and penalizes spurious or chemically implausible assignments. Çubuk says electrical, magnetic or microscopy measurements can help distinguish phases with similar diffraction patterns. Better characterization can make the next experiment more informative; a diffraction match alone does not prove a material is useful.[1]

The team combines human work with automation, choosing routine bottlenecks where machines can help instead of treating full autonomy as the immediate goal. Fedus gives one data-quality example: an AI system noticed measurements that became consistent after reversing a cyclic permutation. The team traced the mismatch to a machine loaded incorrectly. This is one reported diagnosis, not a measured error rate for the lab.[1]

Physical work imposes its own limits. Furnaces drift and measurements are noisy; some superconductivity tests take far longer than diffraction scans. Fedus also describes a planned robot redesign to move an arm’s resting position away from mixed samples and reduce contamination risk. These details matter because a model trained on flawed measurements can learn the wrong lesson.[1]

Periodic says it is adapting tools developed for its own lab to industrial research, with a focus on semiconductors. Fedus describes plans for engineers and researchers to work with partners on site, and says systems can be trained on partner data. The company’s aim is to accelerate researchers’ work; the interview does not provide a quantified customer outcome or demonstrate that it has solved superconductivity.[1]