Descriptors divide‐and‐conquer enables multifaceted and interpretable materials structure–activity relationship analysis
No Thumbnail Available
Date
2025-02-11
Journal Title
Journal ISSN
Volume Title
Publisher
Wiley
Abstract
Machine learning (ML) is increasingly adopted to explore the dependence of properties on descriptors especially for materials with the complicated structure–activity relationships. However, most current ML modeling strategies typically depend on a single combination of descriptors, which leads to inaccurate and unilateral inferences. Here, a descriptors divide‐and‐conquer method is proposed for machine learning (descriptors‐DCML) in which rough set theory (RST) is integrated with materials domain knowledge to select multiple optimal sets of descriptors combinations and thus diverse rule extraction strategies are provided to dig out mechanisms latent in materials data. Its potential utility and applications using the sodium ion energy barrier prediction of NASICION‐type solid‐state electrolyte compounds with multifaceted influencing factors as an example are demonstrated. A total of 85 NASICION‐type samples with 45 descriptors derived from 72 published literature serve as the data foundation for ML modeling. Not only does descriptors‐DCML exhibit the energy barrier prediction accuracy of 93.8% but also extract 9 relations mapping essential factors to Na ion energy barrier in which 5 ones conform to existing understanding rule and the rest are waiting for validation. This work paves the way for reducing the complexity of analyzing materials structure–activity relationships and enhancing the interpretability of ML models. © 1999-2025 John Wiley & Sons, Inc or related companies.
Description
Keywords
Machine Learning, Materials, Sodium ions, Electrolytes, Energy, Sodium, Data, Materials, Accuracy
Citation
Liu, Y., Wu, L., Yang, Z., Zou, X., Zou, Z., Lin, Y., Avdeev, M., & Shi, S. (2025). Descriptors divide‐and‐conquer enables multifaceted and interpretable materials structure–activity relationship analysis. Advanced Functional Materials, 35(26), 2421621. doi:10.1002/adfm.202421621