Comparison of machine learning approaches for the identification of top-performing materials for hydrogen storage

Abstract

Unique properties of Metal-Organic Frameworks (MOFs), such as their extremely high surface areas and porosity, render them as one of, if not the most promising adsorbents for gas storage applications. However, their extremely flexible and tunable nature has resulted in an enormous expansion of the available material pool which in turn begs the question as to whether an efficient identification of the best materials is feasible. In the last years, Machine Learning (ML) techniques have been extensively applied for the exploration of large material databases, since they can significantly accelerate this process. In this work, “traditional” ML models and models based on our recently developed iterative self-consistent approach are compared with respect to their ability to efficiently identify the best materials of a database. As a case study, we have used hydrogen adsorption in MOFs at different thermodynamic conditions. Despite their high accuracy, traditional models struggle to pinpoint the best materials, regarding usable gravimetric uptake, without compromising computational resources. On the other hand, self-consistent models can even reduce by two orders of magnitude the amount of reference data required for the identification of the best gravimetric materials compared to traditional ones. Notably, 300 training samples are enough for the SC models to correctly identify the top-100 gravimetric materials of a database. Nevertheless, both type of models underperform when they are queried for the top-performing materials with regards to usable volumetric uptake.

Keywords

• Machine learning
• Metal-organic frameworks
• Hydrogen storage
• Nanomaterials
• Nanoporous materials
• Gas adsorption