Discovered Materials offers $1M prize pool for AI models that solve five materials science problems

Discovered Materials, a months-old San Francisco startup from Y Combinator's spring 2026 batch, is putting up $1 million in total prizes for AI systems that can crack five long-standing problems in materials science. As TechCrunch reports, the challenges were drawn up with input from an advisory group of academics at leading universities, and are meant to serve as a yardstick for whether AI is genuinely advancing in a field where progress has been hard to measure.

Five problems, one prize pool

Materials science has no equivalent of mathematics' Millennium Prize Problems, so Discovered Materials asked its academic advisers to spell out open questions worth attacking. According to TechCrunch, the five challenges are:

• A theory explaining cuprate superconductivity — why certain materials carry current with zero electrical resistance at ultra-low temperatures.
• A numerical method for predicting the structure of interfaces between silicon and silicon dioxide, a pairing found in nearly every semiconductor.
• A numerical method linking a material's grain structure to its thermal properties, which would help engineers design better heat-dissipating materials.
• A theory for why shining light on a material can change its behavior, particularly around superconductivity.
• A numerical improvement to Density Functional Theory, the workhorse technique researchers use to model atomic and molecular structures.

Felipe Jornada, a Stanford professor who helped develop the list, said in a statement that the field is full of observed phenomena nobody can yet explain, and that explaining them would deepen our grasp of physics while shaping how electronics are built, how energy is stored and how future materials are made.

Frontier models have already tried

Before going public with the competition, Discovered Materials gave frontier lab models a dry run. Per TechCrunch, AI agents equipped with internet access and other resources worked autonomously on each problem for several days — and made little headway. TechCrunch notes it would not be surprising to see a lab such as OpenAI or Anthropic attack the same problems with far more compute or a newer model.

Advaith Sridhar, co-founder of Discovered Materials, told TechCrunch that the hard part is definitional. Mathematics offers crisp yes-or-no criteria for a solution, he explained, while materials science questions are harder to pin down — which is why the professors' help in shaping the right questions mattered so much.

A scoreboard for AI progress

Sridhar framed the prize as deliberately agnostic about who wins. Whether his company, a university researcher or a frontier lab finds the answers, he said, the challenges exist to fix reference points that reveal whether models are improving — in his words, whether they are "climbing in this domain."

He also described the effort as an inversion of the usual scientific workflow: agree on the verification and the challenge first, then let someone — human or machine — supply the ideas that explain the problem.

The startup behind the challenge

Discovered Materials' core business is developing new materials for semiconductors. The company runs research projects with major chip manufacturers while also funding its own scientific work. When it announced its seed round in August, Sridhar told TechCrunch he hoped the company would generate a useful material within a year. He now says those timelines have accelerated, and that he is more confident industrially useful materials will arrive sooner than that.

Why it matters

AI benchmarks in mathematics, coding and games have become familiar, but physical science has lacked an equivalent scoreboard — partly because the problems themselves resist clean, verifiable statement. By working with academics to define five open challenges and inviting anyone to solve them, Discovered Materials is testing whether benchmark-driven AI progress can extend into a domain where the deliverables are theories and numerical methods rather than single answers. The stakes are practical as well as academic: most of the five problems bear directly on chipmaking, heat management or the simulation tools the industry depends on. If models begin closing these gaps, it would be early evidence that AI-driven discovery can operate at the frontier of physics, not just inside well-specified formal systems.