Taylor Sparks1,2,Sterling Baird1,Jason Hall1,3,Ramsey Issa1
University of Utah1,University of Liverpool2,Northrop Grumman Corporation3
Taylor Sparks1,2,Sterling Baird1,Jason Hall1,3,Ramsey Issa1
University of Utah1,University of Liverpool2,Northrop Grumman Corporation3
Real-world materials science optimization tasks are often noisy, heteroskedastic, multi-fidelity, multi-objective, high-dimensional, constrained, and mixed numerical/categorical optimization problems. While each of these have state-of-the-art implementations in computer science, the application of these to materials science tasks have been limited. One of the few optimization platforms that can support all of the use-cases mentioned above without oversimplification and therefore poor efficiency is Meta's Adaptive Experimentation (Ax) platform. While Ax and its backbone, BoTorch, have seen increasing usage in chemistry and materials science, project-specific implementations and adaptations of advanced Bayesian optimization topics are non-trivial, even for veteran materials informatics practitioners. Inspired by the PyTorch installation docs (https://pytorch.org/get-started/locally/), we implement a "choose your own adventure" template generator for Ax scripts tailored towards materials science applications. We demonstrate its usage and performance in three case studies: composition-based optimization, physics-based simulations, and a self-driving lab demo. We envision that this tool will dramatically reduce the barrier-to-entry for utilizing advanced Bayesian optimization for real-world materials science tasks.