Learning from physical experiments
Periodic Labs works on AI systems connected to autonomous laboratories. Its proposed loop is to run physical experiments, collect measurements and use the resulting data to improve models that guide further experiments. Materials science, including superconductors, is a focus. Those are research objectives, rather than evidence that the company has already discovered a commercially usable superconductor. [1]
Its September 2025 public emergence included a $300 million seed round. The technical reason for following the company is the model-and-experiment loop, rather than the financing alone. [3][1]
A model operating inside a laboratory
In September 2026 Periodic described Neon, a trillion-parameter model post-trained for its laboratory work. The account includes X-ray diffraction analysis and learning from data generated by physical experiments. The company reported deploying Neon in its own labs; it did not announce downloadable weights or a generally available model API. [2]
This distinction matters when interpreting AI-for-science progress. An internally deployed model is a concrete research system, while reproducibility and external access are separate questions. The source documents Periodic's methods and claimed results, without independently validating its comparisons against other models. [2]
Sources
- Periodic Labs
Periodic Labs · retrieved Oct 23, 2026
- Building Labs that Learn
Periodic Labs · Sep 15, 2026
- Former OpenAI and DeepMind researchers raise whopping $300M seed to automate science
TechCrunch · Sep 30, 2025