ResearchPapers, evidence & method
Periodic Labs founders detail plan to train AI for materials science with reinforcement learning on their own lab data
Periodic Labs co-founders Liam Fedus and Ekin Doğuş Çubuk say physical experiments, not more text reasoning, limit AI-driven science. Their approach trains agents with reinforcement learning on the company's own lab data, starting with automated analysis of what each synthesis run actually produced.
AT Pulse research deskOctober 9, 2026 · 5 min read · AI-assisted analysisThe company doesn't reward agents for a whole discovery, which would take days of furnace time and give noisy signals. It rewards smaller checkable tasks instead. One example is correctly identifying which crystal phases appear in X-ray diffraction data, with penalties for phases that aren't chemically plausible. To resolve ambiguous patterns, the agents also use electrical, magnetic and microscopy measurements along with chemical priors. This focus on lab characterization follows criticism that earlier AI materials predictions lacked real-world validation. One of those efforts was DeepMind's GNoME, which Çubuk led. [1] [4]
Training tasks are built from Periodic's own experiment records, frozen at set dates, so a model can't recall answers it saw in pre-training. These records include failed syntheses and entire campaigns, data that published papers usually leave out. The founders call full lab autonomy a non-goal. They automate whichever step is currently the bottleneck and see humanoid robots as a slower path. [1]
Periodic launched in 2025 with a $300M seed round led by Andreessen Horowitz. Its early revenue reportedly comes from engineering problems such as heat dissipation for a chipmaker, not from new materials. Later reports describe a round of at least $500M at a valuation of about $7B to $7.5B, and analysts list the pace of discovery relative to spending as a risk. [3] [2] [5] [6] [7]
The evidence so far is early semiconductor revenue and a detailed training design, with no published, validated discovery.
Read the full assessment
If the approach works, records of how experiments were run, including failures, could become the scarcest asset in AI for materials.
Executive brief
Periodic Labs is about a year old and has raised a $300M seed round. It has also been reported in talks at a valuation of $7B or more. Even so, it has published no validated material discovery, and its own site lists no quantified results (periodic.com). In a Latent Space interview, co-founders Liam Fedus (formerly OpenAI) and Ekin Doğuş Çubuk (formerly DeepMind, GNoME) argue that physical experiments, not more reasoning over text, are what limit AI-driven science. Their stated approach is to build reinforcement-learning (RL) environments on top of their own lab data, starting with automated analysis of what each experiment actually produced. The founders described this approach. Outside reviewers have not evaluated it.
What changed and event timeline
Launch with $300M seed
Andreessen Horowitz led the round, with Nvidia, Accel, DST, Jeff Bezos, Eric Schmidt and Jeff Dean also investing. Initial targets were superconductors and heat problems in semiconductors ().
Scrutiny of the field
MIT Technology Review argued that AI materials discovery now has to prove itself in real labs. It cited UC Santa Barbara researchers who found little evidence that DeepMind's GNoME predictions were new, credible and useful materials ().
Early revenue reported
Periodic reportedly had semiconductor customers, revenue (amount not disclosed) and 64 staff. It was also reported to be in talks to raise about $500M at about $7B, which was still unconfirmed as of June ().
Larger round reported
Forbes reported a planned $500M raise (). Forbes Brasil later reported a round of at least $500M at $7.5B, led by AMP ().
Founders' detailed interview
On Latent Space, Fedus and Çubuk described their RL design, how they use lab data, and their plan to run a network of labs ().
Capabilities and access
- No public model, API or version number. Periodic calls itself an "AI research and deployment company" that builds models and autonomous labs (periodic.com).
- Its lab does automated powder synthesis. It sells model access to advanced-manufacturing firms as an "intelligence layer" (Contrary).
- The one customer case on its site is an unnamed chipmaker with heat-dissipation problems, for which Periodic trains custom agents (periodic.com).
Read the full section
- No public model, API or version number. Periodic calls itself an "AI research and deployment company" that builds models and autonomous labs (periodic.com).
- Its lab does automated powder synthesis. It sells model access to advanced-manufacturing firms as an "intelligence layer" (Contrary).
- The one customer case on its site is an unnamed chipmaker with heat-dissipation problems, for which Periodic trains custom agents (periodic.com).
- Its open-source work includes TorchSim, JAX-MD and Materials Project tooling, plus contributions to Megatron (transcript 01:20:07).
Technical analysis for researchers and developers
- Waiting days for a furnace run to say "superconductor or not" is too slow and noisy to train on.
- Guarding against memorized answers. RL tasks are built from the lab's own data, frozen at a given date ("what did the scientist do next?").
- Diffraction is combined with electrical, magnetic and microscopy data and with chemical priors, because different phases can produce the same diffraction pattern (00:23:28).
Read the full section
- Rewards come from small, checkable sub-tasks. Waiting days for a furnace run to say "superconductor or not" is too slow and noisy to train on. Instead, agents are rewarded for correctly identifying which crystal phases are present in X-ray diffraction data, and penalized for implausible phases (00:10:20).
- Guarding against memorized answers. RL tasks are built from the lab's own data, frozen at a given date ("what did the scientist do next?"). Because pre-trained models have never seen this data, they cannot fake the reasoning by recalling answers (00:12:03).
- Several measurements per sample. Diffraction is combined with electrical, magnetic and microscopy data and with chemical priors, because different phases can produce the same diffraction pattern (00:23:28).
- Simulations calibrated to experiments. Density functional theory (DFT, the standard quantum simulation method for materials) is checked against lab results, Materials Project-style. Fedus says only a small fraction of computed candidates ever get made in the lab (00:37:08).
Claims and evidence
- Analyzing what each experiment produced was the main bottleneck, and this work has "shifted significantly" from humans to AI (01:03:22, 01:04:24).
- An AI system spotted that a misloaded machine had shuffled samples in a cycle, by noticing results that were inconsistent across many runs (00:51:40).
- Negative results and whole campaigns, including failed attempts, are used for training (00:56:13, 00:58:13).
Read the full section
- Analyzing what each experiment produced was the main bottleneck, and this work has "shifted significantly" from humans to AI (01:03:22, 01:04:24). This is the founders' description only, with no published metrics.
- An AI system spotted that a misloaded machine had shuffled samples in a cycle, by noticing results that were inconsistent across many runs (00:51:40). This is an anecdote.
- Negative results and whole campaigns, including failed attempts, are used for training (00:56:13, 00:58:13).
- No outside party has corroborated any discovery or benchmark from Periodic.
Context and prior work
- Çubuk led GNoME at DeepMind, and that project's "millions of materials" claim was disputed (MIT TR).
- Çubuk's case for running many trials: nickelates, a near relative of the cuprate superconductors, were eventually shown to superconduct as thin films.
- Competitors include Lila Sciences, Radical AI, Google DeepMind and Microsoft (Contrary).
Read the full section
- Çubuk led GNoME at DeepMind, and that project's "millions of materials" claim was disputed (MIT TR). Periodic's emphasis on synthesis reads as a response to that criticism.
- Çubuk's case for running many trials: nickelates, a near relative of the cuprate superconductors, were eventually shown to superconduct as thin films. He also says the Japanese group that found MgB₂ screened about 30,000 materials (01:21:23).
- Competitors include Lila Sciences, Radical AI, Google DeepMind and Microsoft (Contrary).
Limitations, safety and contested findings
- It cannot capture microstructure, and it struggles with strongly correlated electrons and with predicting superconducting temperatures. Çubuk also doubts that quantum computers would remove these limits (00:15:28, 00:36:39).
- Reports disagree on the founding date: September 2024 (Forbes Brasil), May 2025 (Contrary), and a public launch in September 2025.
- Analysts list as risks the pace of discovery relative to spending, and whether the approach generalizes beyond superconductors (Contrary).
Read the full section
- DFT models perfect crystals. It cannot capture microstructure, and it struggles with strongly correlated electrons and with predicting superconducting temperatures. Çubuk also doubts that quantum computers would remove these limits (00:15:28, 00:36:39).
- Reports disagree on the founding date: September 2024 (Forbes Brasil), May 2025 (Contrary), and a public launch in September 2025. Valuations are reported as $7B and $7.5B.
- Analysts list as risks the pace of discovery relative to spending, and whether the approach generalizes beyond superconductors (Contrary).
Business and practitioner implications
- The competitive advantage is proprietary process data: full records of how experiments were run, including failures. This is data that published papers leave out.
- Periodic earns early revenue from engineering problems such as chip heat, not from discovering new materials.
- Fedus says full lab autonomy is "a non-goal."
Read the full section
- The competitive advantage is proprietary process data: full records of how experiments were run, including failures. This is data that published papers leave out.
- Periodic earns early revenue from engineering problems such as chip heat, not from discovering new materials.
- Fedus says full lab autonomy is "a non-goal." Automation is applied to whatever step is the current bottleneck, and humanoid robots are judged slower (00:50:46).
- For teams building agents on noisy physical data: invest in sensor logging, repeat runs and data-quality checks before scaling up RL.