Periodic Labs Introduces Periodic Neon

Periodic Labs, an AI research and deployment company creating models and autonomous labs to accelerate science, led by Liam Fedus, Dogus Cubuk, and the team, has introduced Periodic Neon, a 1T-parameter model that was post-trained to analyze experimental results and deployed to its physical labs. This is a step towards putting autonomous discovery into the hands of scientists.

Building AI-Powered Materials Labs.

Since launching last year, Periodic has built high-throughput laboratories in Menlo Park and infrastructure for large-scale physics simulations focused on discovering materials such as superconductors and magnets. The company has progressed from its first physical experiments to labs operating 24/7 at significantly higher throughput.

AI is integrated throughout the research process, from analyzing experimental data and identifying issues to deciding which experiments to run next. Periodic uses frontier models where possible and trains proprietary models when external systems fall short or become too costly.

The approach creates a feedback loop: high-throughput experiments generate data for increasingly capable scientific AI, which then guides better experiments. Periodic Neon, an early example, is a trillion-parameter model trained on the company’s lab data that reportedly outperforms GPT-6 Astra on a scientific analysis task while using relatively little compute.

Training Neon Through Scientific Experimentation.

Periodic developed Neon through midtraining and reinforcement learning on data generated by its laboratories. Unlike digital environments, where agents can run millions of fast, verifiable tasks, physical experimentation is constrained by equipment, power, engineering capacity, time, and often ambiguous results.

Periodic addresses this by structuring materials discovery as a continuous three-stage loop: identify promising materials based on predicted stability and properties, determine how to synthesize them, and analyze the resulting materials to verify their structure and performance. The findings then refine the next hypothesis and experiment.

Moving Materials Discovery Toward the Digital Regime.

Periodic aims to reduce the gap between physical experimentation and digital AI training by continuously analyzing existing experimental data rather than leaving compute idle while experiments run.

Neon was trained to interpret X-ray diffraction (XRD) data from Periodic’s labs, a process that has traditionally required significant scientific judgment and analysis. The company has also built scalable infrastructure for training AI models specialized for scientific applications.

The next step is extending AI from analysis to experimentation, directing research campaigns, developing synthesis procedures, and selecting which experiments to run. The long-term goal is to discover materials that do not yet exist and develop the methods to produce them.

𝐑𝐞𝐦𝐨𝐭𝐞 𝐇𝐢𝐫𝐞 𝐨𝐫 𝐁𝐮𝐢𝐥𝐝 𝐖𝐢𝐭𝐡 𝐔𝐬: 𝐁𝐮𝐢𝐥𝐝𝐰𝐞𝐫𝐤𝐬 helps technology companies build in India: two ways. Hire vetted remote engineers, product leaders, designers, and GTM professionals directly onto your team. Or partner with a proven development studio to design, build, and ship your product end-to-end. Learn More At: menlotimes.com/buildwerks