Nanoporous Materials Discovery via Search Bias-Guided Surrogate Modeling

Nanoporous materials (NPMs)--such as MOFs and COFs--offer a path toward cleaner energy and environmental resilience, enabling carbon capture, gas storage, separations, and sensing. However, the pace of discovery remains constrained by expensive and labor-intensive synthesis and characterization processes, which limits iteration and constrains exploration to a narrow slice of a vast design space. This thesis addresses the offline discovery setting where only historical experimental measurements are available, and the goal is to recommend new NPMs that improve upon the best materials observed to date. A common baseline is to train a surrogate model that fits measured properties and then ranks candidates. While effective for prediction, this strategy can be misaligned with discovery: a model optimized to reduce regression error is not necessarily biased toward better-than-seen materials. We propose an optimization-regularized surrogate that augments standard fitting with an explicit optimization bias. The central idea is to learn from the data’s improvement structure: we algorithmically identify and emulate monotonically improving sequences of materials and introduce a regularizer that encourages the surrogate’s score field to be consistent with these improvement directions. This couples predictive accuracy with a principled search bias suited to offline optimization. Across multiple NPM discovery tasks, the proposed surrogate recommends candidates that consistently outperform existing baselines, including regression-only surrogates, one-step and batch Bayesian optimization, and generative-model-based approaches, in best-of-batch outcomes. Overall, this work shows that embedding an optimization bias into surrogate learning materially strengthens offline materials discovery.