Key facts
- Discovered Materials uses AI agents to search for new materials for integrated circuits.
- The startup aims to address the thermal challenges of AI chips.
- Discovered Materials secured $9 million in seed funding from Lightspeed India Partners.
- The company utilizes Anthropic models and custom-trained physics models for material discovery and verification.
- Discovered Materials has already identified several materials with properties similar to those used by major chipmakers.
Discovered Materials is employing AI agents to accelerate the discovery of new materials for integrated circuits, aiming to address the significant heat generation issues in AI hardware. The startup announced it has closed a $9 million seed funding round, led by Lightspeed India Partners, with participation from Peak XV Partners and angel investors.
Co-founders Advaith Sridhar and Akash Ramdas leverage Ramdas's background in materials science and Sridhar's expertise in AI agents. Their platform uses Anthropic models within a custom harness to generate material leads, which are then verified through simulations run by foundational physics models trained by the company. This approach allows for thousands of material guesses per day, a substantial increase from Ramdas's previous rate of about 20 guesses daily during his PhD.
The company has released examples of hundreds of new materials and its "Material Discovery Bench." While competitors like MatNex, SandboxAQ, and CuspAI are pursuing similar goals, Discovered Materials is focusing specifically on the thermal properties of semiconductor materials. They claim to have already identified several materials that match the performance of those used by major chipmakers, though details remain confidential.
A key challenge highlighted by investors is the complexity of finding materials that not only improve thermal performance but are also manufacturable and maintain desirable electrical properties. Hemant Mohapatra of Lightspeed described the process as "playing whack-a-mole with atomic structures," emphasizing that a material's usefulness depends on multiple properties converging. He believes that while predicting novel substances may become commoditized, Discovered Materials' advantage lies in Ramdas's deep expertise and their ability to rapidly experiment and validate candidates in a lab setting.
Discovered Materials plans to patent the use of promising materials in GPUs or their manufacturing processes, licensing them to chipmakers. They anticipate having patentable materials within the next year. However, the article notes that AI-discovered materials have yet to achieve significant commercial impact, with Insilico Medicine's drug candidate being the closest example. The bottleneck, according to Mohapatra, is not in finding candidates but in filtering and synthesizing them effectively. Sridhar acknowledges that the physical process of making materials cannot be significantly sped up, despite the AI-driven discovery phase.