What Actually Matters for Materials Discovery: Pitfalls and Recommendations in Bayesian Optimization

Tristan Cinquin, Stanley Lo, Felix Strieth-Kalthoff, Alan Aspuru-Guzik, Geoff Pleiss, Robert Bamler, Tim G. J. Rudner, Vincent Fortuin, Agustinus Kristiadi

Keywords: Materials discovery, Bayesian optimization, Gaussian process, Bayesian neural network, feature fine-tuning

TL;DR: We systematically investigate the impact of key design choices in Bayesian optimization for materials discovery and identify pitfalls of current methodologies and opportunities for improvements

Abstract: Materials discovery underpins innovation in many fields such as energy storage and therapeutics delivery, but it requires time- and resource-intensive chemical synthesis and testing. Bayesian optimization (BO) with Gaussian processes (GPs) or Bayesian neural networks (BNNs) offers a promising solution to accelerate materials discovery, but the full landscape of features and surrogate models is still poorly understood. In this work, we systematically investigate the impact of key design choices and identify pitfalls of current methodologies and opportunities for improvements. These include: (1) GPs with the default initialization scheme perform \emph{hardly any better} than random policies; (2) BNNs are \emph{highly sensitive} to hyperparameters; (3) expert-designed molecular features \emph{underperform} compared to learned and even simple, generic features; and (4) simple feature fine-tuning significantly enhances performance, contrary to the conventional practice of using fixed molecular features or costly Bayesian fine-tuning schemes. We identify the design choices that actually matter for BO in materials discovery, namely using a simple but well-initialized surrogate model with feature fine-tuning. Our work provides recommendations for practitioners and highlights future research directions toward more cost-effective BO for materials discovery.

Submission Number: 21