Graph neural network-based multi-objective Bayesian optimization for enhanced screening of metal–organic frameworks with optimal separation performance

Abstract

Metal–organic frameworks (MOFs) are porous crystalline materials with applications in gas capture, drug delivery, and molecular separations. While high-throughput computational screening has traditionally identified promising MOFs, recent advances increasingly harness machine learning to accelerate discovery and screening. Existing optimization methods such as Bayesian optimization (BO) and genetic algorithms often overlook the detailed structure–property relationships critical to MOF performance. Here, we present an optimization workflow that couples graph neural networks (GNNs) with multi-objective BO to enhance MOF discovery and screening. By representing MOFs as graphs embedding atomic- and structural-level features, GNNs capture intricate structure–property correlations, enabling more accurate property predictions than traditional methods relying solely on macroscopic descriptors. Our integrated framework efficiently identifies Pareto-optimal MOF candidates tailored for improved separation of alkanes, alkenes, alcohols, and aromatics, demonstrating the significant advantage of graph-based models in materials optimization workflows.