Theory-Guided Machine Learning in Materials Science

Abstract

Materials scientists are increasingly adopting the use of machine learning tools to discover hidden trends in data and make predictions. Applying concepts from data science without foreknowledge of their limitations and the unique qualities of materials data, however, could lead to errant conclusions. The differences that exist between various kinds of experimental and calculated data require careful choices of data processing and machine learning methods. Here, we outline potential pitfalls involved in using machine learning without robust protocols. We address some problems of overfitting to training data using decision trees as an example, rational descriptor selection in the field of perovskites, and preserving physical interpretability in the application of dimensionality reducing techniques such as principal component analysis. We show how proceeding without the guidance of domain knowledge can lead to both quantitatively and qualitatively incorrect predictive models.

Assessing Local Structure Motifs Using Order Parameters for Motif Recognition, Interstitial Identification, and Diffusion Path Characterization

“…3 We ultimately emphasize that materials science is currently undergoing a "change of paradigm: from description to prediction" (Heine, 2014). Thus, we expect these tools to be useful in future machine-learning (Jain et al, 2016;Ward and Wolverton, 2017) applications as descriptors that capture much of the most basic-but essential (Wagner and Rondinelli, 2016)-information of a given material: the crystal structure.…”

A Critical Review of Machine Learning of Energy Materials

“…This down‐selection process can be automatic, e.g., using L 1 or L 0 regularization (least absolute shrinkage and selection operator, LASSO), feature importance, genetic algorithms, etc. However, a drawback of data‐driven feature selection is that the selected features do not imply causality with respect to the target and will be highly dependent on the chosen hyperparameters of the model . Yet another approach is to use dimension reduction algorithms to “synthesize” new and low‐dimensional features from the original features.…”

“…However, a drawback of data-driven feature selection is that the selected features do not imply causality with respect to the target and will be highly dependent on the chosen hyperparameters of the model. [105] Yet another approach is to use dimension reduction algorithms to "synthesize" new and low-dimensional features from the original features. The principal component analysis (PCA) [106] is widely used in this context.…”

Property Predictions for Dual‐Phase Steels Using Persistent Homology and Machine Learning

“…Notably, the persistent diagram was evenly divided into 2500 bins for kernel density estimation in the present work, that is, the source data of PI possesses a dimension of 2500. In machine learning, an excess of model variables often leads to overfitting . Therefore, principal component analysis (PCA) was employed to reduce the dimension of the dataset.…”

“…In machine learning, an excess of model variables often leads to overfitting. [22] Therefore, principal component analysis (PCA) was employed to reduce the dimension of the dataset. PCA normalizes the highdimension dataset with correlated variables and converts it into a set of linearly uncorrelated vectors.…”