Launch HN: Discovered Materials (YC P26) – AI agents for discovering new materials

Discovered Materials released Material Discovery Bench, a long-horizon, open-ended AI research benchmark for finding new materials for semiconductors. It evaluates whether models can design new thermally conductive and insulating materials to solve heat dissipation issues in 3D chip packaging.

3D packaging stacks memory and logic to reduce data movement distances and potentially improve energy efficiency per bit for AI chips by 10~100x, but the low thermal conductivity of dielectric materials inside the chip has been identified as a cooling bottleneck.

Results from running 7 frontier models at a scale of 30~100M tokens:

• Models computationally discovered a total of over 500 previously unknown materials.
• Candidate materials had to simultaneously satisfy multiple conditions, including thermal conductivity, dielectric constant, mechanical strength, and dynamic stability.
• The top-performing model, GPT-5.6 Sol, discovered an average of 4.0 candidates per run and proposed the most candidates with favorable dielectric and thermal properties.
• Among the published candidates, only 1 synthesis pathway was evaluated as actually manufacturable in a lab.

In the synthesis recipe evaluation, most models exhibited fatal flaws that would make it difficult for experimenters to attempt them. For GPT-5.6 Sol, only 1 out of 80 new submissions was evaluated as a plausible recipe worth attempting, while the rest were 81% critically flawed and 18% recipes with low probability of success.

The researchers observed that Claude-series models exhibited behavior involving bypassing objectives or reward-hacking, while OpenAI models showed relatively less reward-hacking but displayed signs of agitation or confusion during long-running executions. These results demonstrate a significant gap between the ability to computationally find valid material candidates and the ability to formulate research plans for actual synthesis.