Explainable machine learning analysis of electrical conductivity in graphene–polymer composites

Abstract

Electrical conductivity in graphene–polymer composites varies over many orders of magnitude because conductive pathways form only when graphene flakes become sufficiently connected. This study analyzes a literature-derived dataset of 486 graphene–polymer composite samples using Random Forest and Gradient Boosting models. The input descriptors were polymer matrix, filler type, processing route, system classification, and graphene loading. The conductivity target was expressed as log₁₀ (s) to reduce its wide dynamic range. On the fixed test split, Gradient Boosting achieved the best performance, with R²=0.91, RMSE=1.36, and MAE=0.93 in log₁₀ (s) space, compared with R²=0.84, RMSE=1.77 and MAE=1.45 for Random Forest. The results indicate that graphene loading alone does not explain conductivity variation across heterogeneous systems. Grouped feature-importance analysis showed that graphene loading (34.1%), system classification (32.6%), and processing route (20.8%) had the largest relative contributions to the model predictions, whereas polymer matrix (9.1%) and filler type (3.2%) showed lower grouped importance. These percentages represent model-based relative importance rather than causal physical contributions. Microstructural descriptors, including graphene aspect ratio, dispersion quality, and agglomeration state, were not consistently reported in the source literature and therefore, they were not included among the five model inputs; their absence remains an important source of predictive uncertainty. Overall, interpretable machine learning provides a useful framework for analysing conductivity trends in literature-derived datasets while making explicit the limitations of the available data.