Benchmark Datasets for Active-Learning-Guided Search-Space Reduction and Multi-Objective Bayesian Optimization

This repository contains the two datasets used to evaluate an active-learning-guided search-space reduction framework for multi-objective Bayesian optimization. The first dataset corresponds to a pressure-vessel design optimization problem and contains 52,272 candidate configurations described by geometric and physical design parameters. The problem is formulated as a three-objective minimization task targeting the reduction of the two principal stress components, (S_{11}) and (S_{22}), and the vessel thickness, (T). It provides a controlled engineering benchmark for evaluating adaptive search-space reduction in a large discrete design space. The second dataset comprises covalent organic frameworks (COFs) and is derived from the computational database reported by Mercado et al. in Chemistry of Materials (2018), which contains 69,840 in silico–assembled COFs evaluated for methane storage applications. The dataset is used here as a representative materials-discovery benchmark for multi-objective optimization. Together, the datasets support the evaluation of a workflow in which active learning is first employed to classify and filter candidate configurations, retaining the more promising region of the design space. Multi-objective Bayesian optimization is subsequently performed on the reduced space, and its performance is compared with Bayesian optimization conducted on the corresponding original design space. The repository provides the tabular datasets used in the experiments reported in the associated manuscript and is intended to facilitate reproducibility and further benchmarking of adaptive search-space reduction methods. *This is an anonymized version of the dataset repository. We prepared this anonymized version only for the double-blind peer review process. The author name will be included upon acceptance from the journal.