Parts | Days | Selection | Search | Updates | Downloads | Help
SYNM: From Physics and Big Data to the Design of Novel Materials
SYNM 1: From Physics and Big Data to the Design of Novel Materials
SYNM 1.2: Invited Talk
Monday, September 5, 2022, 15:30–16:00, H1
Beyond the average error: machine learning for the discovery of novel materials — •Mario Boley1, Simon Teshuva1, Felix Luong1, Lucas Foppa2, and Matthias Scheffler2 — 1Monash University — 2Fritz Haber Institute of the Max Planck Society
Machine learning models promise to radically accelerate the discovery of novel functional materials by rapidly screening huge candidate spaces for materials with rare combinations of properties. While current models allow for accurate property prediction on average, e.g., indicated by the root mean squared error, such measures are only loosely connected to materials discovery. They do not capture that discovery is a process, where models are repeatedly retrained with new data. Moreover, they depend on some fixed sampling distribution, which is irrelevant when actively exploring candidates and distorted by the overwhelming mass of mediocre materials. Finally, they do not reflect model uncertainty, which is a crucial parameter for effective active learning strategies. Here, based on the example of searching for stable and stress resistant double perovskites under a bandgap constraint, we evaluate models in terms of their optimisation regret---a function of consecutive improvements of the target property by newly discovered materials. In particular, we employ non-parametric regression models (Gaussian processes and random forests) within the ``expected improvement'' search strategy. Starting from an initial dataset of 815 computed double perovskite properties, we are able to discover materials with improved target values already within the first 20 newly acquired data points.