Discovered Materials Offers $1 Million for AI Solutions to Five Unsolved Materials Science Problems

Discovered Materials, a San Francisco startup, has announced a challenge offering $1 million in total prizes for solving five materials science problems with AI models. The TechCrunch report dated 5 November 2026 describes the challenge as the company's public test of whether current models can make progress on open scientific questions.

The company uses AI models to search for new substances for building advanced chips. It was part of Y Combinator's spring 2026 batch, and its YC profile lists it as batch P26, founded in 2026, with a team size of two. Advaith Sridhar is a co-founder.

The five problems span theory and numerical methods:

• A theory explaining cuprate superconductivity, which the challenge describes as zero electrical resistance at ultra-low temperatures.
• A numerical method (a computer-based calculation technique) for predicting the structures of interfaces between silicon and silicon dioxide.
• A numerical method that maps grain structure (the arrangement of microscopic crystal regions inside a material) to thermal properties, to support the design of better heat dissipation materials.
• A theory predicting why shining light on materials can change their behavior, particularly around superconductivity.
• A numerical method to improve Density Functional Theory, a widely used computational technique for modeling the structures of atoms and molecules.

An advisory group of academics at leading global universities helped define the five open problems. Stanford professor Felipe Jornada helped develop the challenges.

Before announcing the challenges, Discovered Materials gave frontier AI agents internet access and let them work autonomously for several days on each problem. By the company's account, the agents made little headway. The problems were chosen because general-purpose agents, given time and tools, did not crack them, so a solution would have to come from a method that goes beyond current baseline performance.

In August 2026, Discovered Materials closed a $9 million seed round to fund the hunt for novel materials that could build more efficient chips. Lightspeed led the round. Y Combinator and Peak XV participated, and Paul Graham, Gokul Rajaram and Thariq Shihipar joined as angels. The company said it plans to use the money to scale its AI agents, with the stated aim of compressing years of materials R&D and shortening the lab-to-fab timeline.

The company has also open-sourced Material Discovery Bench, which it describes as a long-horizon, open-ended research benchmark for AI-driven materials discovery. Its tasks ask models to search for new thermally conductive dielectric materials. The company says it built the benchmark with experts from IBM, IMEC, Stanford and Cambridge. Over three months of its Y Combinator batch, according to the company's website, Discovered Materials simulated, synthesized and tested thermal interface materials, and says their performance matched products sold for more than 20 years by the world's largest chemical companies.

Sridhar studied AI at Carnegie Mellon and was a research engineer building video models and agents at Persona AI and Luma Labs, according to the company. The YC profile associates him with IIT Madras. The company's other co-founder, Akash, holds a PhD in materials science from Stanford.

The company's stated focus is the semiconductor industry, specifically datacenters and fabs, and the five problems map onto that focus. Thermal interface materials and dielectrics govern how heat leaves a dense accelerator package. Silicon and silicon dioxide interfaces sit at the heart of transistor behavior. Density Functional Theory underpins much of the computational screening used to propose candidate compounds in the first place.

In my view, a better method on any of these would feed directly into the materials workflow the company is trying to automate, and that is the practical reason the problems were chosen. The more useful question for practitioners is not whether a model wins the $1 million. It is whether the evaluation can be checked. Open problems with no accepted answer are hard to score, and a benchmark is only as good as its judging criteria. The company's decision to publish its agents' near-failures alongside the prize suggests it understands that. Whether the prize pool attracts serious entrants will say more about the field's appetite for hard, verifiable scientific work than about any single model.

The broader context here is a shift in how AI labs and startups frame progress. Benchmarks built on exam questions have largely been saturated, and the appeal of a materials problem is that the answer lives in a physical lab or a validated simulation rather than on a leaderboard. If a method emerges that improves Density Functional Theory or explains a piece of superconductivity, the gain would extend well beyond this competition. That is the upside the challenge is built to test, and its outcome is not yet known.