Evolution-guided Bayesian optimization for constrained multi-objective optimization in self-driving labs

Evolution-guided Bayesian optimization for constrained multi-objective optimization in self-driving labs

2024 | Andre K. Y. Low, Flore Mekki-Berrada, Abhishek Gupta, Aleksandr Ostudin, Jiaxun Xie, Eleonore Vissol-Gaudin, Yee-Fun Lim, Qianxiao Li, Yew Soon Ong, Saif A. Khan & Kedar Hippiagaonkar
The paper introduces Evolution-Guided Bayesian Optimization (EGBO), an algorithm that integrates selection pressure with a q-Noisy Expected Hypervolume Improvement (qNEHVI) optimizer to address constrained multi-objective optimization problems (cMOOPs) in high-throughput experimental platforms. EGBO is designed to efficiently explore the Pareto Front (PF) while avoiding infeasible regions, which is crucial for materials science and chemical engineering applications. The authors demonstrate the effectiveness of EGBO through a self-driving lab for seed-mediated silver nanoparticle synthesis, targeting three objectives (optical properties, reaction rate, and seed usage) with complex constraints. Compared to state-of-the-art qNEHVI, EGBO shows improved performance in hypervolume improvement, PF coverage, and feasible solution generation. The paper also discusses the handling of input and output constraints, proposing a pre-repair method to reduce sampling wastage. Overall, EGBO is presented as a general framework for solving constrained multi-objective problems in high-throughput experimentation platforms, offering a robust and efficient solution to complex optimization challenges.The paper introduces Evolution-Guided Bayesian Optimization (EGBO), an algorithm that integrates selection pressure with a q-Noisy Expected Hypervolume Improvement (qNEHVI) optimizer to address constrained multi-objective optimization problems (cMOOPs) in high-throughput experimental platforms. EGBO is designed to efficiently explore the Pareto Front (PF) while avoiding infeasible regions, which is crucial for materials science and chemical engineering applications. The authors demonstrate the effectiveness of EGBO through a self-driving lab for seed-mediated silver nanoparticle synthesis, targeting three objectives (optical properties, reaction rate, and seed usage) with complex constraints. Compared to state-of-the-art qNEHVI, EGBO shows improved performance in hypervolume improvement, PF coverage, and feasible solution generation. The paper also discusses the handling of input and output constraints, proposing a pre-repair method to reduce sampling wastage. Overall, EGBO is presented as a general framework for solving constrained multi-objective problems in high-throughput experimentation platforms, offering a robust and efficient solution to complex optimization challenges.
Reach us at info@study.space