Active Learning Guided Design Space Refinement for Scalable Multi-Objective Bayesian Optimization in Materials Discovery

As materials discovery datasets grow in size and complexity, conventional Bayesian optimization (BO) often wastes computational resources evaluating low-performing candidates. This paper investigates whether an adaptive, active-learning-based search-space refinement can effectively prune these large candidate pools while preserving the regions necessary to discover high-performing, Pareto-optimal materials.

The researchers propose a two-stage framework. First, they employ a classification-based active learning approach using Density-Aware Greedy Sampling (DAGS) and an XGBoost classifier to partition the design space into promising and non-promising regions. This stage uses a composite proxy score—combining mean, worst-case, ideal-point, and geometric objectives—to label candidates. Second, they construct a reduced candidate space by retaining high-confidence, high-uncertainty, and Pareto-relevant samples. This reduced space is then passed to a multi-objective BO workflow as a warm-start, allowing the optimizer to focus exclusively on the most informative regions.

The framework was tested on two complex materials discovery problems: methane/nitrogen separation in covalent-organic frameworks (COFs) and the design of composite pressure vessels. The results demonstrate that the refinement process successfully reduces the candidate space by approximately 50% while retaining over 99% of the original Pareto-relevant hypervolume. By focusing the BO on these refined subsets, the authors observed improved early convergence and more efficient discovery of the Pareto front compared to standard BO applied to the full, unrefined space.

This approach addresses a critical bottleneck in autonomous materials discovery: the scalability of BO in high-throughput settings. By decoupling the search-space reduction from the optimization process, researchers can handle significantly larger candidate pools without a proportional increase in the number of expensive, high-fidelity simulations required to find optimal material trade-offs.