[Paper Review] Persistent homology advances interpretable machine learning for nanoporous materials.

Aditi S. Krishnapriyan, Joseph Montoya+2 more|arXiv (Cornell University)|Oct 1, 2020

Machine Learning in Materials Science30 references4 citations

TL;DR

This paper proposes using persistent homology to create interpretable, universal structural representations for nanoporous materials, enhancing machine learning models for gas adsorption prediction. By integrating topological features with chemical embeddings, the method improves model accuracy and transferability across diverse targets while enabling pore-level interpretability of structure-property relationships.

ABSTRACT

Machine learning for nanoporous materials design and discovery has emerged as a promising alternative to more time-consuming experiments and simulations. The challenge with this approach is the selection of features that enable universal and interpretable materials representations across multiple prediction tasks. We use persistent homology to construct holistic representations of the materials structure. We show that these representations can also be augmented with other generic features such as word embeddings from natural language processing to capture chemical information. We demonstrate our approach on multiple metal-organic framework datasets by predicting a variety of gas adsorption targets. Our results show considerable improvement in both accuracy and transferability across targets compared to models constructed from commonly used manually curated features. Persistent homology features allow us to locate the pores that correlate best to adsorption at different pressures, contributing to understanding atomic level structure-property relationships for materials design.

Motivation & Objective

• To address the challenge of selecting universal, interpretable features for machine learning in nanoporous materials design.
• To develop a holistic, topologically informed representation of materials structure using persistent homology.
• To improve model performance and transferability across multiple gas adsorption prediction tasks.
• To enable atomic-level interpretability by linking specific pore structures to adsorption behavior at different pressures.

Proposed method

• Employ persistent homology to extract topological invariants from nanoporous material structures, capturing pore connectivity and cavity morphology.
• Construct feature vectors from persistent homology barcodes to represent material topology in a machine learning-ready format.
• Augment topological features with word embeddings from natural language processing to encode chemical composition and elemental information.
• Train supervised machine learning models on metal-organic framework datasets using the hybrid feature representation for gas adsorption prediction.
• Use feature importance analysis to identify pores most correlated with adsorption at varying pressures, enabling interpretability.
• Evaluate model performance and transferability across multiple gas adsorption targets, comparing against models using manually curated features.

Experimental results

Research questions

• RQ1 Can persistent homology generate universal and interpretable structural representations for nanoporous materials across diverse prediction tasks?
• RQ2 How does combining topological features with chemical word embeddings improve model accuracy and generalization?
• RQ3 Which pore structures are most predictive of gas adsorption at different pressures, and can persistent homology identify them?
• RQ4 To what extent does the proposed method outperform models based on manually curated features in terms of accuracy and transferability?
• RQ5 Can persistent homology features enhance interpretability by linking specific topological features to adsorption mechanisms?

Key findings

• The persistent homology-based representation significantly improves prediction accuracy compared to models using manually curated features.
• The method demonstrates superior transferability across multiple gas adsorption targets, indicating robust generalization.
• Topological features enable identification of specific pores most correlated with adsorption at different pressures, enhancing interpretability.
• Augmenting topological features with chemical word embeddings further boosts model performance.
• The approach allows researchers to locate and analyze atomic-level structural motifs responsible for high adsorption capacity.
• The method provides a holistic, interpretable framework for materials design that is both accurate and transferable across diverse prediction tasks.