AQForge: Bridging Generative Models and Property Prediction for Materials Discovery

Abstract

Designing and characterizing new materials with tailored properties is critical in fields such as catalysis, energy storage, and solid-state materials. Despite significant technological advances, including the development of generative models and universal machine learning force fields, these tools often operate in isolation rather than as integrated components of a comprehensive workflow. Additionally, while the computation of energies and forces remains highly valuable and the focus of many studies, it often falls short for accurately predicting certain task-specific macroscopic properties. To address these limitations, we propose an end-to-end workflow that extends the capabilities of current state-of-the-art works and fully automates the design and discovery of materials, with a particular emphasis on calculating downstream properties. By integrating and validating existing approaches, we ensure the robustness of the workflow and demonstrate its utility with a few illustrative use cases.