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

What it claims

Periodic Labs argues that end-to-end RL on "did we find a room-temperature superconductor" is infeasible. Each rollout takes days of furnace and instrument time, and there are too few agents to reduce variance. Instead they decompose the materials discovery loop into sub-tasks, each with its own tractable RL environment trained on data from their own labs. The interview gives two concrete environment designs and several practical reasons why simulation (DFT) cannot replace the lab.

How it works

The loop. The loop has three stages:

1. Predict what to make: a stable configuration that also has the target property.
2. Find synthesis conditions.
3. Characterize the output, which "doesn't come out of the furnace labeled."

A large volume of DFT calculations is filtered, by humans or LLMs or both, down to what actually gets run in the lab. DFT is calibrated against experiment per chemical system, following the approach Materials Project used.

Environment 1: phase identification from XRD.

• The model gets raw X-ray diffraction data.
• It is rewarded for correctly identifying the phases present.
• It is penalized for spurious phases or chemically implausible ones.
• Real outputs are usually mixed: leftover precursors, amorphous material, unintended phases, and sometimes phases recorded in no database.
• Ambiguity remains because two phases can fit the same pattern. They resolve it with:
• replicates
• thermodynamic and synthesis-condition priors
• multimodal measurements (electrical, magnetic, electron microscopy) that separate phases XRD cannot
• This work also removes a throughput bottleneck: at high experiment rates, interpreting what was made becomes the limit.

Environment 2: timestamped evidence.

• Snapshot the internal experimental record as of a given date.
• Ask the model for the scientist's next choice, or for the outcome of the next experiment.
• The stated rationale: on public data, a pretrained model may already have memorized the answer. It can then "fake work," skipping the physics and simulation, and RL reinforces reasoning that will not generalize to novel systems. Unpublished in-house data with full lineage avoids this.

Noise sources they model or instrument against:

• furnace heating elements degrading with each use as material evaporates onto them
• non-uniform furnace temperature by position
• actual temperature differing from logged temperature
• differences between machines
• vibration affecting optical setups

They respond with heavy telemetry and replicates, to cut down hidden variables. They state they never delete data.

Evidence

The page gives no benchmarks, accuracy figures, throughput numbers or model details. All claims are qualitative statements from the founders.

Limits

• Where DFT falls short:
• It typically simulates perfect crystals, so it misses microstructure.
• It fails under strong electron correlation.
• It cannot easily predict properties such as superconducting critical temperature.
• Cubuk says even exact DFT would not remove the need for a lab, and is skeptical that quantum computing changes this: you would "still [be] simulating perfect crystal."
• XRD is an average. It is not a unique structure: disproportionation can hide behind a perfect-looking average pattern. Inorganic crystals also lack a SMILES-like 1D representation.
• Sample efficiency is the binding constraint. Unlike digital environments, you cannot add rollouts arbitrarily.
• Truncated page. The fetched text covers only about the first 46 minutes of an 85-minute transcript. It does not include the sections on:
• "140 IQ" instruments
• lab automation
• negative results
• training on the process of science
• scaling labs
• industry deployment
• the superconductor search

Claims from those sections are not reflected here.