SBIR Phase I: AI-Driven Cloud Application Programming Interface for Quantum-Accurate Materials Simulation

Award ID: 2603825

The broader/commercial impact of this Small Business Innovation Research (SBIR) Phase I project is the development of an AI-driven platform that accelerates discovery of advanced materials critical to next-generation electronics, including semiconductors, energy systems, quantum computing, and advanced manufacturing. Today, identifying and testing new materials often requires hundreds of thousands to millions of high-performance compute hours, taking months to complete. This project develops technology designed to reduce these calculations to minutes, lowering computational costs by more than 90% and enabling researchers to evaluate hundreds of thousands of candidate materials in weeks rather than months. This capability is increasingly important for U.S. national security and technological competitiveness. Advanced materials underpin next-generation semiconductor chips, AI computing infrastructure, defense systems, and energy technologies, yet first-principles simulation remains a major bottleneck limiting commercialization and deployment. This technology is designed to remove this bottleneck and accelerate the pace of American scientific and industrial innovation. This Small Business Innovation Research (SBIR) Phase I project aims to solve a severe computational bottleneck in materials science: the exponential cost of quantum mechanical simulations required for next-generation electronics discovery. The research objective is to develop a foundational AI model that predicts the electronic structure of solid materials with high-fidelity accuracy in seconds. To achieve this, the proposed research involves constructing a massive, rigorously standardized training dataset derived from plane-wave calculations. Using this uniform data, the project will train an equivariant graph neural network that natively outputs the effective Hamiltonian in a maximally localized, completely orthogonal Wannier basis. Unlike competing machine learning tools that produce non-orthogonal outputs requiring computationally expensive matrix inversions, this orthogonal architecture uniquely enables direct subspace projection. By isolating specific energy windows near the Fermi level, this capability dramatically reduces the Hilbert space, turning previously intractable downstream transport and scattering computations into routine tasks. The anticipated technical result is a linear-scaling computational framework capable of evaluating systems containing over one million atoms. Ultimately, this technology will provide the fast, quantum-accurate engine needed to accelerate the autonomous evaluation and design of novel semiconductors, quantum processors, and advanced energy storage materials. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

NSF Program Director: Vincent Lee