Description
Following the transformative impact of AlphaFold generative models are emerging as a major opportunity in materials science. Among these candidates, frameworks such as GNOME andMatterGen promise to generate vast libraries of materials, yet the primary bottleneck remains the rapid and reliable screening of the enormous number of structures they produce. DECLARE directly addresses this challenge. As a key demonstration, we focus on metallic alloys. Despite centuries of research into these essential structural materials, predicting and controlling the thermodynamic phases of metallic alloys remains a grand challenge in computational materials science. By providing a robust, high-throughput link between atomic-scale descriptors and phase stability, DECLARE offers the critical screening capability that will make the next generation of generative materials design truly predictive.
Recent advances in statistics, artificial intelligence and machine learning (AI/ML) have now made tractable to address directly phase stability of materials. This progress stems from extensive community efforts improving data-driven force fields for interatomic interactions and developing foundational models, as well as making free energy methods more efficient compared to earlier approaches. By integrating free energy calculations with machine learning potentials into streamlined workflows, constructing “phase diagrams” has become feasible, though not yet routine and with huge numerical effort. The DECLARE project goes beyond these advances by fundamentally reframing the problem through the innovative concept of the descriptor’sdensity (DDOS, described below).
The main outcome will be a validated predictive phase diagram workflow that represents a major advance in materials design by replacing empirical rules with quantitative, actionable data. We propose to (i) develop and distill advanced machine learning foundation models capable of accurately and efficiently predicting atomic-scale properties, leveraging entropy descriptor-based free energy methods (DDOS); (ii) create a comprehensive thermodynamic database integrating free energies and configurational entropy across the periodic table, employing Bayesian adaptive sampling to capture metastable states and refine alloy phase stability predictions with DDOS; (iii) integrate these theoretical and mathematical findings with the experimental existing information into a unified predictive framework that constructs validated phase diagrams, quantifies uncertainties, and enables rapid, generalized materials design under kinetic and thermodynamic constraints.The D-DOS method, the core of this IRC, is a game-changing approach for free-energy calculations. Several of the key ideas behind this method were made possible thanks to the participation of TDS and MCM in a long program at the Institute for Mathematical and Statistical Innovation (IMSI), University of Chicago. (NSF grant DMS-1929348).
Dates
June 8-18, 2026