Machine Learning for Discovery of Novel Quantum Materials: From Data-Driven Screening to Physics-Constrained Closed-Loop Discovery

Machine learning (ML) is revolutionizing the discovery of quantum materials by addressing the immense chemical and structural search space and the prohibitive computational cost of first-principles density functional theory (DFT). This review provides a comprehensive examination of modern ML paradigms, ranging from crystal graph neural networks and equivariant interatomic potentials to active learning, Bayesian optimization, generative diffusion models, and emerging multimodal foundation models. We critically analyze the integration of these techniques with high-throughput DFT and experimental workflows. A central thesis of this work is that high predictive accuracy on benchmark datasets is a necessary but insufficient condition for genuine scientific discovery; a computationally stable candidate may be dynamically unstable, experimentally inaccessible, or lie perilously outside the model’s training distribution. Consequently, we advocate for a paradigm shift from static screening toward physics-constrained closed-loop discovery systems. Such systems must seamlessly integrate generative candidate design, uncertainty-aware property prediction, selective high- fidelity quantum verification (including DFT+U, GW, and DMFT), and iterative experimental feedback. Recent landmark achievements, including the GNoME expansion of stable crystals, the development of universal potentials like M3GNet and CHGNet, the conditional generation capabilities of MatterGen, and closed-loop experimental discovery of superconductors, exemplify this transformative transition. The remaining grand challenge is not merely scaling model parameters, but architecting scientifically trustworthy systems that inherently encode physical symmetries, reliably quantify epistemic uncertainty, distinguish interpolation from extrapolation, and continuously evolve from sparse, expensive quantum mechanical calculations and experimental measurements to accelerate the discovery of novel, synthesizable quantum materials with targeted electronic, magnetic, and topological properties.

کلیدواژه ها:

Quantum materials ، Machine learning ، Active learning ، Generative models ، Materials discovery.