Novelty-Aware Evolutionary Bayesian Optimisation for Multi-Objective Discovery Science

Abstract

Efficient optimisation of complex experimental systems is a central challenge in modern discovery science, particularly in settings characterised by high-dimensional design spaces, expensive evaluations, and multiple competing objectives. Multiobjective Bayesian optimisation (MOBO) has emerged as a leading approach for such problems due to its sample efficiency, but can suffer from limited exploration and reduced diversity, especially in many-objective, multimodal, and constrained settings. Evolutionary algorithms, by contrast, excel at maintaining diversity across the Pareto front but typically require large evaluation budgets. Here, we systematically investigate hybrid evolutionary-Bayesian optimisation strategies that combine the strengths of both approaches. Building on the Evolutionary Guided Bayesian Optimisation (EGBO) framework, we benchmark multiple evolutionary generators within a unified acquisition-driven pipeline across ten synthetic test problems spanning multimodal, many-objective, and constrained regimes. We further introduce a novelty-aware batch selection strategy that explicitly promotes diversity within candidate batches while retaining model-guided prioritisation. Across benchmarks, hybrid methods consistently outperform acquisition-only MOBO in challenging optimisation regimes, achieving improved hypervolume, lower inverted generational distance, and more reliable convergence. Gains are most pronounced in many-objective and multimodal problems, as well as in feasibility-limited search spaces. However, performance advantages diminish in very high-dimensional feature spaces, where evolutionary exploration reduces sample efficiency. The proposed novelty-aware selection further improves performance by reducing redundancy within batches and mitigating optimisation stagnation. Importantly, these trends translate to real-world experimental datasets spanning reaction optimisation, pharmaceutical formulation, materials design, and drug screening. Together, these results demonstrate that hybrid evolutionary-Bayesian optimisation provides a robust and practical strategy for improving optimisation performance in autonomous and data-driven discovery workflows. Introduction