Radical AI

Radical AI develops materials through self-driving laboratories: AI generates hypotheses, robots synthesize and characterize samples, and experimental results feed back into learning. Its approach connects computational predictions to physical testing while running research campaigns in parallel. Joseph Krause, a materials scientist and cofounder, frames the requirement plainly: “In materials, the ground truth is the material itself. You have to be able to test it and characterize it.”[1][10][5]

The laboratory matters because a material's composition alone does not determine its performance. Mixing, annealing, and other manufacturing processes can produce different microstructures from the same chemical ingredients. Radical's integrated system therefore links a material's digital blueprint, physical creation, and measured performance rather than treating simulation and experimentation as separate stages.[1][4]

That integration strategy took public shape in 2025 through a simulation-software release, an Air Force research award, and seed financing, following the company's founding in 2024.[7][6][4][5]

AlleyWatch reports that Joseph Krause and Jorge Colindres founded Radical AI in 2024. On July 18, 2025, Radical announced the close of a $55 million seed round; AlleyWatch identified RTX Ventures as the lead investor. Radical described its funding strategy as building materials intelligence that combines scientific literature, computational models, and self-driving-lab data in parallel.[5][4]

Radical also received a $1,197,902 Phase II USAF SBIR award for autonomous experimentation on materials for hypersonic development. The official award record names Krause as principal investigator and gives a contract schedule of May 21, 2025, through February 23, 2027. The proposed work combines computational screening, AI models, and robotic laboratories to design and test high-entropy alloys. The amount is a research-contract award, separate from the seed financing.[6]

In his 2025 recap, Krause reported that Radical had grown to 36 people, built six custom laboratory automation stations, and synthesized and characterized more than 700 novel alloys. He also reported signing a lease for a 45,000-square-foot headquarters. Antimatter, the company's LabOS, shipped versions V2, V3, and V4 during the year, according to his account.[12]

Radical released TorchSim on April 4, 2025, as a PyTorch-native atomistic simulation engine. It batches mathematical operations across multiple systems, allowing them to evolve concurrently on GPUs. The framework supports molecular dynamics, structural relaxation, and both machine-learned and classical interatomic potentials. Radical chose to open-source this software layer while stating that its commercial strategy would be to sell materials to energy, aerospace, defense, semiconductor, and automotive industries.[7][9]

The TorchSim paper, submitted on August 8, 2025, credits Orion Cohen, Janosh Riebesell, Rhys Goodall, Adeesh Kolluru, Stefano Falletta, Joseph Krause, Jorge Colindres, Gerbrand Ceder, and Abhijeet S. Gangan.[9]

Radical's next public research contribution addressed how AI interprets experiments. Its February 10, 2026, MATRIX announcement distinguished MATRIX, a multimodal materials-science reasoning benchmark, from MATRIX-PT, its corresponding post-training dataset; a model was released separately. Tasks cover foundational theory, research-level reasoning, hypothesis generation, and interpretation of microscopy images, diffraction patterns, elemental maps and spectra, and thermal curves.[10]

The MATRIX paper, submitted January 30, 2026, was authored by Delia McGrath, Curtis Chong, Rohil Kulkarni, Gerbrand Ceder, and Adeesh Kolluru. The researchers compared post-training on structured materials-science text with training that also used paired experimental images. They reported improvements in experimental interpretation and text-only scientific reasoning that depended on correct image–text alignment, along with gains on ScienceQA and PubMedQA. This work evaluates scientific reasoning rather than laboratory throughput.[11][10]

Brandon Anderson's June 17, 2026, Latent Space episode editorial reported that Radical had produced and characterized 1,200 alloys in six months. It also reported an AI-scientist campaign that proposed and tested 300 materials, ten of which had novel state-of-the-art properties and were undergoing further development for commercial applications.[1]

On August 7, 2026, Radical reported a campaign to replace aerospace material C103, generating more than 300 novel materials over 16 weeks. Working with Purdue Applied Research Institute, it compared candidate RAI-939 with C103 in a one-minute torch test. Radical reported a 0.25% mass gain for RAI-939 and a 32% mass loss for C103. Its stated next step was scaling alloy production to make larger, manufacturing-representative test articles and gather quality-control data for a specification sheet.[14]

Sample preparation exposed a practical constraint on laboratory automation: button-sized alloy samples varied in size, requiring repeated reconfiguration of an electrical discharge machining (EDM) machine before cutting pieces for different tests. In its August 26, 2026, engineering account, Radical described working with Makino to address the bottleneck. Mechanical engineer Nathan Kadria and his team developed a metal insert that cuts five samples simultaneously and deposits the pieces into separate baskets. Radical reported that the insert made cuts on delivery day. Makino sales engineer David Lovejoy credited Radical's engineers with drawing on Makino's expertise while asking detailed technical questions.[13]