Berlin 2024 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 11: Data Driven Material Science: Big Data and Workflows II
MM 11.4: Talk
Monday, March 18, 2024, 16:30–16:45, C 243
From Prediction to Action: Critical Role of Performance Estimation for Machine-Learning-Driven Materials Discovery — •Lucas Foppa1, Mario Boley2, Felix Luong2, Simon Teshuva2, Daniel Schmidt2, and Matthias Scheffler2 — 1The NOMAD Laboratory at the FHI of the MPG and IRIS-Adlershof of the HU Berlin, Germany — 2Department of Data Science and AI, Monash University, Australia
The development of machine-learning models for materials properties focuses on improving the average predictive performance of the models with respect to some training-data distribution. However, a good performance in average might not translate into an efficient discovery of materials via model-driven blackbox optimization (e.g., Bayesian). In these iterative materials-discovery approaches, the training data is extended based on a model-informed acquisition function whose goal is to maximize a cumulative reward over iterations, such as the maximum property value discovered so far. Crucially, the rewards might be decoupled from the average predictive performance, as they can be dictated by the model performance for the few exceptional materials of interest. Here, we illustrate this problem for the example of bulk-modulus maximization in perovskites and propose an estimator that recovers qualitative aspects of the actual rewards and can be computed using the intial training data.[1]
[1] M. Boley, et al., arXiv:2311.15549 (2023).
Keywords: Black-Box Optimization; Bayesian Optimization; Materials Discovery; Artificial Intelligence