Accelerated development of multi-component alloys in discrete design space using Bayesian multi-objective optimisation

Abstract Bayesian optimisation (BO) protocols grounded in active learning (AL) principles have gained significant recognition for their ability to efficiently optimize black-box objective functions. This capability is critical for advancing autonomous and high-throughput materials design and discovery processes. However, the application of these protocols in materials science, particularly in the design of novel alloys with multiple targeted properties, remains constrained by computational complexity and the absence of reliable and robust acquisition functions for multiobjective optimisation. Recent advancements have demonstrated that expected hypervolume-based geometrical acquisition functions outperform other multiobjective optimisation algorithms, such as Thompson Sampling Efficient Multiobjective optimisation and pareto efficient global optimisation (parEGO), in both performance and speed. This study evaluates several leading multiobjective BO acquisition functions–namely, parallel expected hypervolume improvement (qEHVI), noisy qEHVI, parallel parEGO, and parallel noisy parEGO (qNparEGO)–in optimizing the physical properties of multi-component alloys. Our findings highlight the superior performance of the qEHVI acquisition function in identifying the optimal Pareto front across 1-, 2-, and 3-objective aluminum alloy optimisation problems, all within a constrained evaluation budget and reasonable computational cost. Furthermore, we explore the impact of various surrogate model optimisation methods from both computational cost and efficiency perspectives. Finally, we demonstrate the effectiveness of a pool-based AL protocol in expediting the discovery process by executing multiple computational and experimental campaigns in each iteration. This approach is particularly advantageous for deployment in massively parallel high-throughput synthesis facilities and advanced computing architectures.