Structural Feature Extraction via Topological Data Analysis

With the advancement of artificial intelligence models, the development of scientifically grounded and structurally appropriate feature extraction methods has become critical for machine learning-based structure prediction and materials design. In recent years, there has been growing dissatisfaction with inefficient empirical descriptors and black-box feature extraction processes that require extensive training. This article introduces a topological data analysis-based framework for extracting structural features of materials, offering an informative perspective on structure-property relationships and predictive strategies. Emphasis is placed on the predictive power and interpretability of topological features, highlighting their advantages in uncovering structure-property correlations and providing physical insights into material behavior. This approach establishes a mathematically rigorous and computationally efficient paradigm for the discovery and design of advanced materials, achieving up to 55% reduction in prediction error for defect-sensitive properties and a notable improvement in MOF gas uptake prediction accuracy (e.g., R² from 0.74 to 0.85), thus demonstrating both theoretical clarity and practical performance.
© 2025 American Chemical Society