Session:

The development of Materials Acceleration Platforms has enabled High-Throughput (HT) materials experimentation towards solving materials grand challenges with large decision spaces, multiple conflicting objectives and complex constraints. The implementation of existing state-of-the-art Bayesian Optimization called q-Noisy Expected Hypervolume Improvement (qNEHVI) is able to solve noisy constrained multi-objective problems with batch evaluations, appropriate for HT experimentation. However, the stochastic sampling approach used to generate candidates leads to inefficient exploration of the search space and inability to fully resolve the Pareto Front, reducing the number of good solutions that can be tested experimentally. In this work, we develop a new algorithm specifically designed for materials discovery using a hybrid framework called Evolution-Guided Bayesian Optimization (EGBO) that introduces selection pressure via evolution to rapidly navigate the exploration-exploitation trade-off. We demonstrate this on multiple multi-objective benchmark problems, both synthetic and real-world, and show that EGBO is able to efficiently resolve the Pareto Front [1]. Asides from hypervolume, we present other metrics that are specifically relevant to assessing optimization for material science problems such as scalability, consistency and feasibility. We demonstrate this on a closed-loop experimental workflow that explores Cu-Sb-S compositional space to optimise its photoelectrocatalytic hydrogen evolution from water,. The multiple-objective optimization model discovered optimum experimental conditions after only nine cycles of integrated experiments-machine learning loop. Photocurrent testing at 0 V vs. reversible hydrogen electrode (RHE) confirmed the expected correlation between the materials properties and photocurrent. An optimum photocurrent was observed on Cu₉Sb₄₅S₄₆ in the form of single layer coating on F-doped SnO₂ (FTO) glass with corresponding bandgap of 1.85 eV and 63.2% Cu¹⁺ species content [2].

[1] EGBO: Evolution-guided Bayesian Optimization for Efficient Pareto Front Exploitation for Materials Discovery, submitted to Nature Computational Science (2022)
[2] Multi-Objective Optimization for Photoelectrocatalytic Materials Discovery via High-throughput Experiments Combined with Machine Learning, submitted to Advanced Materials (2022)

Co-Author(s): Kai Yuan Andre Low (Nanyang Technological University), Yang Bai (Institute of Materials Research and Engineering), Khoo Zi Hui Jonathan (Institute of Materials Research and Engineering), Made Riko I (Institute of Materials Research and Engineering), Eleonore Vissol-Gaudin (Nanyang Technological University), Huiqing Xie (Institute of Materials Research and Engineering), Albertus D. Handoko (Institute of Materials Research and Engineering), Vijila Chellappan (Institute of Materials Research and Engineering), Jianwei Jayce Cheng (Institute of Materials Research and Engineering), Fengxia Wei (Institute of Materials Research and Engineering), Yee Fun Lim (Institute of Materials Research and Engineering)

Kedar Hippalgaonkar

Nanyang Technological University

Appearances