Periodic Labs’ AI Scientists Need the Physical Experiment

What came out of the furnace?

A materials experiment begins with a recipe and ends with a sample, not a label. Mix precursor crystals, heat them, and the result may contain some starting material, an amorphous component and several unintended crystalline phases. Before deciding what to try next, a scientist has to work out what was actually made.

In a Latent Space conversation, Periodic Labs’ Liam Fedus and Ekin Doğuş Çubuk describe building AI systems around that discovery loop. Fedus separates the work into choosing a stable candidate with a desired property, finding a way to synthesize it, and characterizing the resulting sample. Each step can fail independently: a material that looks promising on paper may be difficult to make, and the product may not be the intended one.

The laboratory is essential to their thesis. Fedus argues that ideas and simulations must be tested against physical outcomes. He says a direct reinforcement-learning loop that waits days for each experiment and rewards only the discovery of a room-temperature superconductor would be too slow and noisy to provide enough useful feedback.

A fingerprint with more than one explanation

One useful intermediate task is phase identification. Fedus describes shining X-rays at a sample: atomic spacings produce a diffraction pattern that acts like a structural fingerprint. Periodic’s early AI work, he says, includes systems that fit the observed data, reward identifying phases present and penalize implausible or spurious phases. Better analysis matters because high-throughput experiments can outpace the people interpreting them.

The fingerprint is not always unique. Fedus says different phases can fit the same pattern, so the system needs chemical context and repeated measurements. Çubuk adds other views: electrical or magnetic properties and electron microscopy can distinguish candidates that look similar in X-ray diffraction. These signals reduce ambiguity without giving a perfect view of every atom.

Çubuk frames the deeper problem as incomplete observation. A physical system contains far more atomic detail than can be observed or stored; useful variables such as temperature or pressure compress it. Even the variables that are recorded can mislead: he notes that furnace position can affect temperature and a heating element can degrade over time. More telemetry helps, but it does not reveal the complete state of the experiment.

Simulate, make, compare

Simulation helps decide which experiments to run. Çubuk describes using density functional theory, or DFT, to estimate formation enthalpy and assess whether a proposed material is energetically favorable relative to alternatives. DFT uses electron density to study a quantum system; he distinguishes the underlying theory from the approximations needed in practical calculations.

A calculation is not the same thing as a fabricated sample. Çubuk says simulations commonly describe ideal crystals, while real materials have microstructure and imperfections that can affect their properties. Some properties, including superconducting temperature, are difficult to simulate with DFT. He says Periodic calibrates simulations against experimental data in chemical systems of interest.

This makes the loop practical: use calculations and prior knowledge to filter candidates, synthesize selected ones, characterize the result, and use those observations to revise the next choice. Fedus and Çubuk describe this as their approach and ambition; the interview does not present a controlled end-to-end discovery benchmark or establish that the loop has solved superconductivity.

The data includes the attempts that failed

A null result can mean the intended structure was not produced. Fedus says it can also reveal an unexpected structure instead. Çubuk stresses a crucial qualification: failure under one recipe is evidence about that attempt, not proof that the material can never be synthesized. Published literature tends to emphasize successful crystals, leaving useful negative examples underrepresented.

A sequence of failures followed by success can teach more than the final recipe alone. Fedus wants to retain the choices, computations, conversations and lab operations across a campaign so models can learn the process of iteration. He also proposes training tasks from a time-stamped record: give a model the evidence available at a past point and ask about the subsequent decision or result.

The distinction guards against a misleading success: if a model has memorized a published answer, it can arrive there without doing reasoning that transfers to a new material. Fedus proposes privately generated experimental evidence as one way to construct tasks less exposed to that shortcut. Reusing the record in multiple training environments expands the number of learning tasks, not the number of independent physical measurements.

Make instruments more useful

Fedus gives a concrete automation example from scanning electron microscopy. A simple program captured a view and zoomed in, but its images were not sufficiently useful for the scientists’ question. He says Periodic built AI control that takes the experiment’s intent and other evidence into account while choosing what to capture. The goal is a richer record at the time of measurement.

Even in experiments lasting hours or days, individual control or analysis steps can be latency-sensitive. Fedus notes that repeated model and tool calls can add up; Çubuk points to the cost and difference between waiting minutes and hours for an X-ray analysis. The right inference arrangement depends on the physical process and the decision being made.

Automation also serves data quality. Fedus recounts an incorrectly loaded machine that created a cyclic mismatch in measurements; an AI system noticed that reversing a permutation made the records consistent, and the team traced the issue to loading. As a hardware-design example, he says a robotic arm resting over previously mixed samples posed contamination risk, so the team planned to redesign it with the resting position away from those samples.

The objective is useful, diverse, high-quality data rather than full autonomy for its own sake, Fedus says. A routine operation may be worth automating while a difficult dexterity task remains with a person. He describes campaign guidance as increasingly mixed between people and AI and characterization as shifting toward AI assistance; the exact balance changes with the work.

More chances to discover

At the end of the conversation, Çubuk distinguishes proposing a promising chemical family from actually synthesizing it. He points to the eventual fabrication of superconducting nickelates in thin-film form as an illustration of why a candidate idea and the process to realize it can be separated by years of work. He presents faster, broader experimentation as a way to increase the opportunities for discovery—not as a demonstrated route to a particular breakthrough.

Periodic is also taking its tools into industry: Fedus says the company is building a forward-deployed effort with a focus on semiconductors, where teams would integrate and adapt systems within partners’ secure environments. He describes outcome-based pricing as a possible future as autonomy improves. Those plans and company-reported capabilities should be evaluated separately from a verified materials discovery.