Topological descriptor for interpretable thermal transport prediction in amorphous graphene

Understanding and predicting thermal transport in disordered materials remains a significant challenge due to the absence of periodicity and the complex nature of medium-range structural motifs. In this work, we investigate amorphous graphene and demonstrate that persistent homology, a topological data analysis technique, can serve as a physically interpretable structural descriptor for predicting thermal conductivity. We first show that ridge regression using persistent homology descriptors achieves high prediction accuracy. To gain physical insight into the prediction process, we perform inverse analysis by mapping the regression coefficients back onto the persistence diagrams. This reveals that distorted hexagonal and triangular motifs are strongly correlated with reduced thermal conductivity. A further comparison with the spatial distribution of localized vibrational modes supports the physical interpretation that these motifs suppress thermal transport. Our findings highlight that persistent homology not only enables accurate physical property prediction but also uncovers meaningful structure – property relationships in two-dimensional amorphous materials. This approach offers a promising framework for interpretable machine-learning models in materials science.