Transferable and Robust Machine Learning Model for Predicting Stability of Si Anodes for Multivalent Cation Batteries

Abstract

Data-driven methodology has become a key tool in computationally predicting material properties. Currently, these techniques are priced high due to computational requirements for generating sufficient training data for high-precision machine learning models. In this study, we present a Support Vector Regression (SVR)-based machine learning model to predict the stability of silicon (Si) - alkaline metal alloys, with a strong emphasis on the transferability of the model to new silicon alloys with different electronic configurations and structures. We elaborate on the role of the structural descriptor in imparting transferability to the model that is trained on limited data (~750 Si alloys) derived from the Material Project database. Three popular descriptors, namely X-Ray Diffraction (XRD), Sine Coulomb Matrix (SCM), and Orbital Field Matrix (OFM), are evaluated for representing Si alloys. The material structures are represented by descriptors in the SVR model, coupled with hyperparameter tuning techniques like Grid Search CV and Bayesian Optimization (BO), to find the best performing model for predicting total energy, formation energy and packing fraction of the Si alloy systems. The models are trained on Si alloys with lithium (Li), sodium (Na), potassium (K), magnesium (Mg), calcium (Ca), and aluminum (Al) metals, where Si-Na and Si-Al systems are used as test structures. Our results show that XRD, an experimentally derived characterization of structures, performs most reliably as a descriptor for total energy prediction of new Si alloys. The study demonstrates that by qualitatively selection of training data, using hyperparameter tuning methods, and employing appropriate structural descriptors, the data requirements for robust and accurate ML models can be reduced.

arXiv comment23 pages, 7 figures

References in corpus (18)

• Gaussian Approximation Potentials: the accuracy of quantum mechanics, without the electrons
• Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties
• On representing chemical environments
• Fast and Accurate Modeling of Molecular Atomization Energies with Machine Learning
• AFLOW: An automatic framework for high-throughput materials discovery
• Moment Tensor Potentials: a class of systematically improvable interatomic potentials
• DScribe: Library of Descriptors for Machine Learning in Materials Science
• Machine learning prediction errors better than DFT accuracy
• A critical examination of compound stability predictions from machine-learned formation energies
• Orbital Graph Convolutional Neural Network for Material Property Prediction
• Gaussian approximation potential modeling of lithium intercalation in carbon nanostructures
• Comparative Computational Study of the Energetics of Li, Na, and Mg Storage in Amorphous and Crystalline Silicon
• Systematic beyond-DFT study of binary transition metal oxides
• A Perspective on Deep Learning for Molecular Modeling and Simulations
• MT-CGCNN: Integrating Crystal Graph Convolutional Neural Network with Multitask Learning for Material Property Prediction
• Neural Networks Potential from the Bispectrum Component: A Case Study on Crystalline Silicon
• Packing Fractions and Maximum Angles of Stability of Granular Materials
• Discovering Superhard B-N-O Compounds by Iterative Machine Learning and Evolutionary Structure Predictions