Prediction and optimization of transverse thermal conductivity of green fiber composites based on interpretable machine learning

Kesheng Cai, Chenkai Zhu, Haoran Bai, Xiaoyu Zhao, Guannan Wang,
Prediction and optimization of transverse thermal conductivity of green fiber composites based on interpretable machine learning,
Materials Today Communications,
Volume 49,
2025,
113863,
ISSN 2352-4928,
https://doi.org/10.1016/j.mtcomm.2025.113863.
(https://www.sciencedirect.com/science/article/pii/S235249282502375X)
Abstract: Unidirectional green fiber composites (UGFCs), with outstanding thermophysical properties and sustainability, have become extensively employed in thermal management devices. The application of UGFCs demands accurate prediction of thermal behaviors and efficient composite design. To address the shortcomings from analytical and numerical simulations, this paper adopts interpretable machine learning (ML) methods to predict the transverse thermal conductivity of UGFCs and conduct inverse optimization based on engineering needs. Latin hypercube sampling (LHS) generated 4850 data points, which were then used to train and test ML models (RFR、XGBR、SVR、DNN and CNN). The SVR demonstrated superior performance, achieving an R2 greater than 0.99 on the testing set, with MAE and RMSE values of 0.00125 and 0.00164, respectively. Compared with the theoretical model and the finite element model in new in-domain data, SVR not only maintained a mean relative error of just 0.12 %–0.45 % but also saved 2–3 orders of magnitude in computational time. In addition, it exhibited satisfactory extrapolation capability. Shapley analysis visualizes the influence of input on the output. The results show that the matrix thermal conductivity (Km) has the largest impact, and the fiber volume fraction (Vfb) ranks second, which provides interpretability to the model’s predictions. The Shapley results and SVR are combined with the sparrow search algorithm (SSA) to obtain the minimum transverse cross-sectional area under given component thermal conductivities. This study provides a novel approach for the exact performance prediction and efficient design of the transverse cross-section of unidirectional green fiber composites.
Keywords: Unidirectional green fiber composites; Transverse thermal conductivity; Interpretable machine learning; Sparrow search algorithm; Transverse cross-sectional optimization