Periodic Labs builds AI scientists and autonomous laboratories to accelerate scientific discovery in the physical sciences. The company combines frontier AI models with real-world experimental data, focusing on closing the loop between hypothesis generation and physical reality through reinforcement learning environments. Their autonomous laboratories generate gigabytes of experimental data that feeds back into model training, addressing the fundamental bottleneck of grounding AI systems in physical constraints rather than purely digital reasoning.
The technical architecture spans the full stack: training infrastructure using Megatron-LM, DeepSpeed, FSDP, and TorchTitan; inference deployment with vLLM and SGLang; and physical simulation through COMSOL and ANSYS. The autonomous lab infrastructure handles robotics control, CAD integration, and CUDA-accelerated computation. The team runs weekly teaching sessions where physicists train LLMs on quantum mechanics reasoning while ML researchers learn physics fundamentals - a bidirectional knowledge transfer that directly shapes model capabilities and experimental design.
Target verticals include semiconductors (heat dissipation optimization), superconductor discovery, space, and defense applications where experimental iteration cycles are expensive and domain expertise is scarce. The team comprises physicists, chemists, and ML researchers operating with minimal boundaries between disciplines - cross-functional ownership extends from model architecture decisions through physical lab automation design. The operational model emphasizes rapid experimentation: hypothesis generation by frontier models, automated physical validation, data ingestion back into training loops, and iterative refinement of both model capabilities and laboratory automation.