Liam Fedus

Liam Fedus is a machine-learning researcher and co-founder of Periodic Labs, where he is applying AI to the messy feedback of physical experiments. In a materials lab, a sample comes out of a furnace without a label. Fedus describes using X-ray diffraction patterns to identify the phases that formed, so that a lab can learn from the experiment and choose what to try next.[1][2]

Before Periodic, Fedus co-authored the 2021 Switch Transformer paper with Barret Zoph and Noam Shazeer. Its mixture-of-experts approach routes an input to selected parameters rather than running every parameter for every input. The authors simplified routing to make sparse models easier to train while addressing communication cost and instability. The result is a way to increase model capacity without using the full model for every computation.[3][2]

Fedus and Ekin Doğuş Çubuk built Periodic around a different challenge: a model can propose an experiment, but the physical world determines the outcome. Fedus argues that improving a hypothesis requires running the test, interpreting its measurement and updating the next choice. Instruments drift, temperatures vary, and measurements can be noisy or incomplete; real experiments also take time and cannot be multiplied as freely as digital simulations.[1][2]

A materials search has several linked jobs. A proposed arrangement of atoms must be stable and have a useful property; researchers must find conditions to synthesize it; then they must characterize what actually formed. Fedus describes X-ray diffraction as a practical bottleneck: the pattern acts as a fingerprint for crystalline phases, but identifying a mixture is difficult. Periodic’s early AI work has targeted phase identification, rewarding correct assignments and penalizing spurious or chemically implausible ones. That is a narrower task than discovering a superconductor outright.[1]

He also describes keeping the sequence of experiments and measurements together. By training on the evidence available at a particular point and asking what came next, researchers can test decisions against previously unseen outcomes. Fedus cautions that a model that has memorized the answer may appear to reason without doing the work needed to generalize to a new system.[1]

The lab’s hardware shapes the data as much as its models do. Fedus describes a planned redesign after the resting position of a robotic arm raised a contamination risk above mixed samples. He says Periodic is taking its computational tools to industry, with a focus on semiconductors, and plans to place engineers and researchers on site. He describes training systems on partners’ data as a capability. These are company-reported work and plans; they do not establish a measured improvement for every customer.[1]