AI-Driven Quantum Simulations for Materials Discovery: A Graph Neural Network and Active Learning Framework to Accelerate DFT-based Screening

Abstract

The discovery of novel materials with tailored electronic, mechanical, and thermodynamic properties is a longstanding challenge in materials science, often hindered by the computational expense of high-throughput experimentation and density functional theory (DFT) simulations. This work presents an AI-driven framework that integrates graph neural networks (GNNs), active learning, and generative modeling with quantum simulations to accelerate materials discovery. Crystals are represented as graphs, enabling GNNs to predict key properties such as formation energy and band gap while estimating uncertainty. An iterative active learning loop selects high-potential or high-uncertainty candidates for DFT validation, improving prediction accuracy while reducing computational costs. Generative models propose novel crystal structures satisfying target property constraints, expanding the discovery space. Evaluated on 1,000 candidate structures, the framework demonstrated progressive improvement in prediction accuracy from 85% to 94% over five iterations, with mean DFT error decreasing from 0.25 eV/atom to 0.12 eV/atom. This closed-loop approach not only enhances screening efficiency and enables the identification of high-performing materials absent from existing datasets, but it also accelerates the development of advanced materials for energy storage, electronics, and healthcare applications, ultimately contributing to technological innovation and societal well-being.