Radical AI has produced and characterized 1,200 alloys in six months, nearly 10x the output of the DARPA/GE MACH program, which targeted 500 new alloys in a full year. Founder Joseph Krause built the company around a single hard truth: no AI model can one-shot a material. A chemical formula is not a material. The manufacturing process, microstructure, and synthesis conditions determine what you actually get, and the LK-99 debacle of 2023 proved that gap is not theoretical.
Radical's answer is a closed-loop self-driving lab where an AI scientist generates hypotheses and robotic systems synthesize and characterize outputs in parallel. From one research campaign, the system proposed 300 new materials and identified 10 with novel state-of-the-art properties now moving toward commercial development. The AI has also expanded into elemental and alloy families with no prior published research, which matters not just scientifically but for supply chain resilience in aerospace, defense, and computing. Krause's position is direct: the experimental data is the moat, not the model.
The original conversation with Krause covers the mechanics of the SDL architecture, his time in Washington and his read on China's lab-to-manufacturing pipeline, and what national-level investment in self-driving lab infrastructure would actually require. Radical has also open-sourced key internal tooling including TorchSim, a PyTorch-based molecular dynamics simulation framework now operating as a nonprofit, and the MATRIX foundation model for materials. The open-source releases alone are worth reading for anyone working at the intersection of ML and physical science.
[READ ORIGINAL →]