Olympus, enhanced: benchmarking mixed-parameter and multi-objective optimization in chemistry and materials science

Abstract

Experiment planning algorithms are a required component of autonomous platforms for scientific discovery. Selecting a suitable optimization algorithm for a novel application is an important yet difficult choice a researcher has to make based on past empirical performance on similar tasks. To facilitate the evaluation of various algorithms on chemistry and materials science optimization tasks, we previously introduced OLYMPUS (Mach. Learn.: Sci. Technol. 2, 035021, 2021), a Python package providing a consistent and easy-to-use interface to numerous optimization algorithms and benchmark datasets. While the original package was limited to continuous parameters and single objectives, in this work we expand OLYMPUS' capabilities to include mixed (continuous, discrete, and categorical) parameter types and multiple objectives. Several experiment planning algorithms already contained in OLYMPUS are extended to handle categorical and discrete parameter types, and five additional planners are implemented (23 in total). We also provide 23 additional benchmark datasets taken from the chemistry and materials science literature (33 in total), covering a wide range of research areas, from chemical reaction optimization to materials manufacturing. Finally, the visualization capabilities of OLYMPUS are enhanced to allow for easy inspection of the results, and the core functionality of the package is embedded in a Streamlit web application for code-free usage. We demonstrate how OLYMPUS enables researchers to rapidly benchmark different optimization strategies and gain insight into their behavior by focusing on two case studies: the optimization of a Suzuki-Miyaura cross-coupling reaction with categorical reaction conditions, and the multi-objective optimization of redox-active materials. The updated OLYMPUS package provides practitioners with a large suite of tools to efficiently benchmark and analyze experiment planning algorithms on mixed-parameter and multi-objective optimization tasks.

Discrete and mixed-variable experimental design with surrogate-based approach

“…Previous studies have proposed methodologies to address this challenge, for example, one can integrate surrogate models into these optimization frameworks, which has shown promise in improving convergence rates and solution quality 20,[42][43][44] . Specifically, several optimization algorithms have been developed within the BO framework 40,[45][46][47][48][49] .…”

“…For instance, for fully-categorical design space, the Hamming distance kernel is commonly used. For mixed-integer cases within the reaction optimization domain, it is recommended to use Matérn5/2 kernel for both continuous and integer (discrete) variables 4,20 . Similar to other BO-based approaches, the solving process includes two stages -the initial and activelearning stages.…”

“…A reaction scheme for the investigated Suzuki-Miyaura coupling is shown in Fig. 4, illustrating the coupling of a boronic acid derivative and an aryl halide facilitated by a palladium complex catalyst, a ligand, a base, and a solvent 4,20,71 . Here, all optimization variables are categorical and the number of possible options for each optimization variable is summarized in Table 1.…”

“…The second case study explores the optimization of the design of a crossed barrel (see Fig. 7) for improved mechanical properties 20,35 . Specifically, we aim to maximize its toughness while not exceeding a specified force threshold.…”

“…Recently, more flexible and adaptive data-driven approaches, such as Bayesian optimization (BO) methods 4,[16][17][18] , have been proposed as a solution, making it possible to adjust the experimental planning dynamically based on accumulated information. For instance, different surrogate-based optimization methods [19][20][21] have been developed over the years. By incorporating Bayesian principles, the experimental space can be navigated more effectively, allowing for more comprehensive exploration and a better chance of uncovering hidden insights that might be overlooked by conventional DoE approaches.…”