Unified AI Framework for Accelerated Materials Design and Discovery

The discovery and design of advanced materials are essential for technological innovation. However, navigating the vast combinatorial space of possible materials and optimizing their properties remains slow and inefficient. To address this challenge, we develop an AI framework that integrates generative models, large language models, and foundation models for atomistic simulations to automate the discovery of materials with targeted properties. Leveraging curated datasets, we train generative and large language models for inverse design, enabling the generation of candidate structures with desired characteristics. For forward evaluation, predictive models are trained to predict a wide range of material properties, including mechanical, thermal, electronic, and dielectric properties. Efficient atomistic simulations are achieved using TorchSim, a next-generation molecular dynamics engine designed for GPUs. This AI loop is managed by autonomous agents that generate, evaluate, and submit candidate materials to our lab, enabling continuous, self-directed exploration of the materials space. By integrating inverse design with predictive evaluation within this agent-driven framework, we efficiently identify novel compositions with optimal property combinations. We demonstrate this approach on bulk metallic glasses, high-entropy alloys, and dielectric materials, highlighting both the novelty of the generated structures and the underlying chemical and structural trends. This framework provides a scalable, systematic pathway for accelerated, automated materials discovery.