A Hybrid Machine Learning Framework for Predicting Hydrogen Storage Capacities in Metal Hydrides: Unsupervised Feature Learning with Deep Neural Networks

TL;DR

This paper introduces a hybrid machine learning framework combining unsupervised autoencoder feature extraction with a deep neural network to accurately predict hydrogen storage capacities in metal hydrides, addressing data scarcity and complexity.

Contribution

The study presents a novel hybrid approach that leverages autoencoders and deep neural networks for improved prediction of hydrogen storage, including transfer learning strategies and material discovery insights.

Findings

Autoencoder-derived features enhance prediction accuracy.

Transfer learning from autoencoder weights slightly improves model performance.

The framework shows good agreement with density functional theory results.

Abstract

In this study, we present a sophisticated hybrid machine-learning framework that significantly improves the accuracy of predicting hydrogen storage capacities in metal hydrides. This is a critical challenge due to the scarcity of experimental data and the complexity of high-dimensional feature spaces. Our approach employs the power of unsupervised learning through the use of a state-of-the-art autoencoder. This autoencoder is trained on elemental descriptors obtained from Mendeleev software, enabling the extraction of a meaningful and lower dimensional latent space from the input data. This latent representation serves as the basis for our deep multi-layer perceptron (MLP) model, which consists of five layers and shows good precision in predicting hydrogen storage capacities. Furthermore, our results show very good agreement with the results of density functional theory (DFT). In addition to addressing the limitations caused by limited and unevenly distributed data in the field of hydrogen storage materials, we also focus on discovering new materials that show promising opportunities for hydrogen storage. These materials were identified using both feature-based approaches and predictions generated by a large language model. Finally, our investigation into the effectiveness of transferring weights from the autoencoder to the MLP, in addition to the latent features, suggests that while this strategy slightly improves model performance indicated by a slightly higher R$^2$ value and lower RMSE, it emphasizes the intricate challenge of adapting pre-trained weights for specific supervised tasks.