RAMBOAU: Risk-averse Multiobjective Bayesian Optimization with Aleatoric Uncertainty for Nanomaterial Synthesis

Optimizing the synthesis conditions of advanced materials is challenging, especially when outcomes are subject to inherent experimental uncertainties. Bayesian optimization is a popular tool for accelerating materials discovery, but its standard risk-neutral framework overlooks the variability of outcomes under different synthesis conditions. This can be misaligned with the need for reliable synthesis protocols that consistently produce materials with desired properties. This work introduces a risk-averse multi-objective Bayesian optimization algorithm that accounts for heteroscedastic experimental uncertainty in the outcomes of different synthesis conditions using a value-at-risk approach. It employs two Gaussian processes per objective, one for the expected function value and one for its variance, and optimizes a novel risk-averse acquisition function. Benchmarking against a state-of-the-art risk-neutral algorithm shows superior performance in identifying robust Pareto fronts. Applied to optimize the production of high-purity, high-aspect-ratio ZnO nanoparticles in a continuously operated annular flow microreactor in the laboratory, the algorithm identified reliable process conditions despite significant uncertainty. This work represents an important step towards autonomous platforms that can identify synthesis conditions that are robust enough for industrial scale-up and practical implementation.