A Hybrid Machine Learning Framework for Predicting Hydrogen Storage Capacities in Metal Hydrides: Unsupervised...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Predicting hydrogen storage capacity in metal hydrides is challenging due to limited experimental data and complex feature spaces.
• Accurate prediction is crucial for developing efficient hydrogen storage solutions.

Purpose of the Study:

• To develop a hybrid machine learning framework for accurate prediction of hydrogen storage capacities.
• To discover novel metal hydride materials for hydrogen storage applications.

Main Methods:

• Utilized unsupervised learning with an autoencoder trained on elemental descriptors to create a lower-dimensional latent space.
• Employed a five-layer deep multilayer perceptron (MLP) model for capacity prediction.
• Integrated a fine-tuned GPT-2 large language model (LLM) for materials generation.

Main Results:

• The hybrid ML framework achieved high accuracy in predicting hydrogen storage capacities, showing agreement with DFT calculations.
• Identified new potential hydrogen storage materials using feature-based approaches and LLM predictions.
• Validated a subset of newly discovered materials using DFT calculations.

Conclusions:

• The developed hybrid ML framework effectively addresses data scarcity and complexity in predicting hydrogen storage.
• The study successfully discovered novel hydrogen storage materials, demonstrating the potential of integrated ML and LLM approaches.