Oct 8, 2026 · 1h 24m

Periodic Labs puts experiments inside the AI reasoning loop

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

The episode examines whether scientific AI can discover useful materials without confronting the slow, noisy, incomplete evidence of physical experiments.

3 key takeaways

1Periodic combines models, simulations, instruments, and laboratories because intelligence alone cannot establish how materials behave in reality.
2Materials discovery depends on modeling synthesis, microstructure, phase transitions, and uncertainty—not merely predicting idealized atomic structures.
3Capturing failed experiments, instrument telemetry, and scientific lineage could give future agents better training data than published results alone.

Don't miss

The guests argue that future scientific agents need the full lineage of discovery, including failed experiments and intermediate reasoning, not just final papers.

The brief

Liam Mendoj and Ekin Doğuş Çubuk present Periodic’s central thesis: scientific intelligence must be grounded in laboratories, where experiments supply noisy evidence that models cannot generate alone.

The materials-discovery loop runs from selecting a promising configuration to synthesizing and characterizing it. Simulations narrow the search, but physical results calibrate models and expose what idealized structures miss.

Phase transitions, microstructure, and incomplete measurements make discovery an inference problem under uncertainty. Periodic combines physics-based priors, replicates, multimodal characterization, and instrument telemetry to reduce ambiguity.

The most consequential proposal is to capture the whole scientific process—conversations, code, simulations, failed syntheses, and intermediate decisions—rather than training agents only on polished publications.

Periodic’s ambition extends from automated instruments and distributed laboratories to semiconductor manufacturing, with humans, targeted automation, and specialized hardware sharing the work.