Radical AI’s self-driving lab claims 10x speedup in alloy discovery with closed-loop AI scientist

Materials discovery has long suffered from slow, serial processes that can span decades from lab to commercial use. Radical AI argues that the bottleneck is not the model but the lab infrastructure and the experimental data it generates. The company’s approach centers on a self-driving lab that integrates an AI scientist with automated robotics to run closed-loop research campaigns, generating, synthesizing, and characterizing materials in parallel rather than sequentially.

Radical AI reports that its self-driving lab produced and characterized 1,200 alloys in six months, a throughput the company says is nearly 10 times faster than the DARPA/GE MACH program, which aimed to create 500 new alloys in a year. The lab’s AI scientist proposed 300 new material candidates, with 10 demonstrating novel state-of-the-art properties and already advancing toward commercial applications. Joseph Krause, founder of Radical AI, emphasizes that experimental data—not a single model—is the true competitive moat in materials science.

The self-driving lab operates as a closed-loop system: the AI scientist generates hypotheses, robotic systems synthesize samples, and characterization tools measure outcomes, feeding results back into the AI for iterative refinement. Radical AI claims it can scale further, estimating the capacity to produce and test a hundred new alloys in a single day. The company also reports that its AI scientist has explored elemental families previously overlooked by human researchers, a development the company frames as both scientifically novel and strategically important for reducing supply-chain vulnerabilities.

Beyond internal R&D, Radical AI has open-sourced parts of its tooling stack. This includes TorchSim, a PyTorch-based molecular dynamics simulation framework released as a non-profit, and MATRIX/MATRIX-PT, an open dataset and model for benchmarking autonomous self-driving labs. The company notes that improvements in reasoning for materials science have unexpectedly translated to gains in biological systems, underscoring the generality of the approach.