Item

Abstract

Materials discovery has traditionally been guided by empirical approaches, with workflows that involve much trial and error. This method has yielded great discoveries of novel materials in the past centuries with the technology available, however, with the development of computational power as well as new methods for developing predictive models for material simulation, approaches are being developed that seek to involve less trial and error, saving time and resources.
Machine Learning (ML) models for various materials properties have become increasingly popular because of their potential to rapidly screen materials at a computational cost orders of magnitude smaller than traditional ab initio methods. However, the current ML methods only exploit the outcome of the model developed and focus on optimizing this model, neglecting the possibility of guiding the next best experiments.
Active learning is the approach of adaptively guiding the choice of the next best experiments, allowing for the best choice of materials to test next in a materials discovery workflow. Despite the promise of active learning to accelerate materials discovery, it is currently still a field that has yet to be explored thoroughly. This provides us with the impetus to explore the active learning method and workflow to materials discovery, and develop our own active workflow and test the feasibility of this approach.
Hence, we propose to develop an active learning scheme by firstly identifying a global model for predicting the bulk modulus of selected perovskite materials, using an AI-based method known as the sure-independence screening and sparsifying operator (SISSO) method. Next, we identify significant areas of interest in our dataset using subgroup discovery (SGD) to find the domain of applicability (DA), and lastly apply the subgroups found to attempt to direct our next chosen area of study.