Discovered Materials offers a million dollars for five unsolved AI research challenges
Discovered Materials has announced five open materials-science challenges and a combined million-dollar prize pool. The startup wants to test whether artificial intelligence can make verifiable progress on longstanding research problems, including superconductivity and semiconductor interfaces. No solution or discovery accompanies the announcement. Proposed advances must meet the requirements set for each scientific problem.
Artificial Intelligence··Night
Open problems replace familiar test questions
Discovered Materials, a startup working on artificial intelligence for materials research, has announced five open scientific challenges with a combined prize pool of one million dollars. Each solved problem carries two hundred thousand dollars. The initiative asks models to tackle questions researchers have worked on for years, rather than answer familiar benchmark questions. The problems remain open at the launch. The San Francisco company was founded by Advaith Sridhar and Akash Ramdas, both alumni of the Indian Institute of Technology Madras, and joined Y Combinator’s spring cohort. Its work uses groups of AI agents to search for chip materials.[1], [2]
Challenges span superconductors and interfaces
One challenge seeks a predictive theory of cuprate superconductivity. Another asks for the electrical properties of semiconductor interfaces to be predicted from their manufacturing history. A further problem concerns improving the trade-off between calculation speed and accuracy in quantum chemistry. The remaining questions involve calculating how light changes a real material and predicting heat transport across solid interfaces. Each concerns a different part of the same announced research programme. The individual problem pages set out the required content of submissions and link to the participation rules. The proposed work must therefore satisfy a particular scientific question, beyond producing a fluent explanation of the subject.[1]
Academics help define the challenges
Stanford materials researcher Felipe Jornada and other academics helped frame the challenges so that a proposed answer could be evaluated. Sridhar described the difficulty of making an open physical-science question precise: unlike a mathematical problem, it may not begin with a definition leading directly to a yes-or-no answer. Before the public launch, the startup let frontier models with internet access work autonomously for days on the questions. Sridhar told TechCrunch that these trials made no meaningful progress. The company then opened the tasks to outside participants. Judges will have the question-specific submission requirements and participation rules when considering an entry.[1], [2]
