ACED: Accelerating Materials Discovery by Learning with Physics-Informed Constraints

About This Grant

This project aims to revolutionize how scientists discover new materials, which are essential for advancing technologies like solar cells, batteries, medicine, and catalysts. Materials discovery is often driven by complex and expensive simulation methods, typically accurate but time-intensive. The emerging machine and deep learning (MDL) models, trained on the massive simulated data collected over the past decades, offer a faster alternative. Unfortunately, these mdoels often fail to predict critical material properties, such as material stability, that the current project shall focus on. This project addresses these shortcomings of MDL models by integrating fundamental physical principles into the training of MDL models. By doing so, the models will better capture the intricate relationships between materials, leading to more reliable predictions. The project will benefit science and our society in a multitude of ways. Faster and more accurate materials discovery will accelerate innovation in clean energy, electronics, healthcare, national security, and more. The project also contributes to the advancement of artificial intelligence by introducing new techniques for constrained learning, which will impact fields beyond materials science. Additionally, the project emphasizes education and engagement by providing interdisciplinary training for graduate students, equipping them with skills at the intersection of computer science, materials science, and engineering. The goal of this project is to boost the capability of deep learning models to accelerate the discovery of solid-state inorganic materials, based on the massive publicly available simulated data from the gold-standard density functional theory computation based on quantum mechanics. The central hypothesis is that the current Deep learning models suffer in stability prediction because materials are presumed independent during training, against the physical fact that stability is intrinsically a property defined by material groups. To test this hypothesis and remedy this deficiency, this project will introduce a new approach to training models for materials discovery by incorporating explicit and implicit physics-informed constraints to encode the thermodynamic stability of materials with respect to phase transition and phase separation . The effect of these constraints will be assessed using existing benchmark problems and by generating additional data based on quantum mechanics to assess tailored problems in the spaces of random structure searching and fine-tuning universal machine learning interatomic potentials. The proposed work can lead to new developments not only for materials discovery but also for constrained stochastic optimization in numerical optimization and structured-output modeling in deep learning. 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.