Collaborative Research: CDS&E: Motif-Based Deep Learning for Discovery and Chemical Insights into Disordered Phases

Nontechnical Description: Modern technologies, from energy-efficient electronics to batteries, depend on discovering new materials with tailored properties. Today, powerful computer models help identify promising materials before they are synthesized, greatly accelerating scientific discovery. This project develops a machine learning framework that explicitly learns and predicts atomic disorder in crystalline materials. Rather than treating materials as collections of individual atoms, the approach recognizes recurring local structural building blocks, or “motifs,” that capture major materials characteristics such as chemical interactions and material stability. Using an important family of compounds as a model system, the project links quantum mechanical simulations, materials databases, and deep learning to uncover the chemical principles that determine when ordered or disordered structures form. Beyond improving the reliability of computational materials discovery, the project produces open-source software and datasets, develops educational resources at the intersection of materials science and machine learning, and trains students in computational materials design and machine learning. By enabling more accurate and interpretable prediction of complex materials, the research strengthens the nation's capabilities in advanced materials discovery for future energy and electronic technologies. Technical Description: Many technologically important materials exhibit finite-temperature site disorder that strongly influences their stability and functional properties. This project develops a motif-based deep learning framework for predicting and interpreting structural disorder in crystalline solids, using Zintl phases as a representative model system. The central hypothesis is that chemically meaningful local motifs, rather than individual atomic sites, constitute the appropriate representation for learning the energetics of disorder and uncovering its underlying chemical origins. The project constructs a comprehensive database of ordered and disordered Zintl phases through large-scale data mining of crystal structure repositories and first-principles density functional theory calculations. Chemically informed motif representations is derived from recurring coordination environments and anionic building blocks and integrated into a motif-based graph neural network capable of predicting formation energies, structural stability, and order-disorder behavior without direct first-principles calculations. The framework s differentiable motif matching, uncertainty quantification, and active learning to guide targeted density functional theory calculations and experimental validation of high-value predictions. By learning interpretable motif-level energetics, the model reveals the chemical interactions governing disorder formation and establish transferable design principles for complex materials. The resulting computational framework will be able to provide a new paradigm for interpretable machine learning of disordered crystalline materials and accelerate the discovery of functional compounds for thermoelectrics, spintronics, solid-state batteries, and related energy technologies. At a broader level, the project disseminates open-source software and FAIR datasets while providing interdisciplinary training in materials informatics, and first-principles materials modeling. 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.