Auto-MatRegressor: liberating machine learning alchemists.

ABSTRACT

Machine learning (ML) is widely used to uncover structure-property relationships of materials due to its ability to quickly find potential data patterns and make accurate predictions. However, like alchemists, materials scientists are plagued by time-consuming and labor-intensive experiments to build high-accuracy ML models. Here, we propose an automatic modeling method based on meta-learning for materials property prediction named Auto-MatRegressor, which automates algorithm selection and hyperparameter optimization by learning from previous modeling experience, i.e., meta-data on historical datasets. The meta-data used in this work consists of 27 meta-features that characterize the datasets and the prediction performances of 18 algorithms commonly used in materials science. To recommend optimal algorithms, a collaborative meta-learning method embedded with domain knowledge quantified by a materials categories tree is designed. Experiments on 60 datasets show that compared with the traditional modeling method from scratch, Auto-MatRegressor automatically selects appropriate algorithms at lower computational cost, which accelerates constructing ML models with good prediction accuracy. Auto-MatRegressor supports dynamic expansion of meta-data with the increase of the number of materials datasets and other required algorithms and can be applied to any ML materials discovery and design task.