Optimization problems are ubiquitous across all disciplines of chemistry: finding the lowest energy state of a molecular system, determining conditions for chemical reactions or identifying materials with desired properties. In the past years, optimization procedures assisted chemists in finding candidates for organic photovoltaics, predicting reaction paths, and automated experimentation. However, optimization problems in chemistry can involve a large number of parameters, and their evaluations are costly, as they involve experiments or demanding computations.<br/>Recently, we proposed Phoenics for an efficient search of these large parameter spaces. This universal deep Bayesian optimization algorithm is gradient-free and was designed for the optimization of objectives that are expensive to evaluate. By combining concepts from Bayesian optimization and ideas from Bayesian kernel density estimation, Phoenics shows desirable computational properties, such as linear scaling with the dimensionality of the parameter space and number of observations. It employs an intuitive sampling strategy, balancing exploration and exploitation for an efficient parallelization of the global search of the optimum. This sampling strategy also significantly improves the performance of the optimization procedure by simultaneously proposing multiple parameter points sampled with exploration or exploitation biases. Recent extensions to Phoenics include the simultaneous optimization of multiple objectives, as well as the support of continuous, discrete and categorical parameters.<br/>In this talk, I will spotlight the implementation of Phoenics and demonstrate how it is used to find answers to challenging problems in chemistry. In particular, I will highlight applications to the autonomous calibration of robotic hardware, the optimization of reaction conditions targeting multiple objectives, the design of candidate molecules for organic flow batteries, as well as the inverse design of superconducting circuits for quantum computing.<br/>With its flexibility and broad applicability, Phoenics is a step forward towards solving the design challenges of modern chemistry.

Phoenics: A Bayesian algorithm for optimization and inverse design

256th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.