Abstract: With the rapid development of electronic devices, energy storage systems, and aerospace technologies, the demand for polymer-based composites with high thermal conductivity (TC) has become increasingly urgent. However, traditional trial-and-error research approaches heavily depended on individual research experience, and extensive experimental validation often fails to reveal the complex relationships between composite formulations and thermal performances. Herein, an efficient machine learning (ML)-based framework for screening composites with high TC was proposed, aiming to optimize the material design process, reduce experimental costs, and identify the key factors influencing TC. Firstly, a comprehensive dataset comprising 370 valid samples is constructed from published literature for thermally conductive carbon-filled/polymer composites. Subsequently, two hybrid models of Logistic+XGBoost and Logistic+fully connected neural network (FCNN) (LFCNN) are developed via a residual-corrected strategy by integrating Logistic regression with XGBoost or a FCNN to improve the capacity of capturing both linear and nonlinear features. Noticeably, LFCNN exhibits the best performance with a classification accuracy of 96.77%. Quantitative analysis using SHapley Additive exPlanations (SHAP) reveals that numerical features exhibit significantly higher importance than categorical features, with filler loading identified as the dominant factor. Within a high-throughput designed compositional space of 32327 sample groups, three promising candidates were selected using LFCNN and experimental validation, where the measured TC is fully consistent with the model predictions. This ML-based framework demonstrates the great promise of integrating ML with experimental validation to accelerate the development of high-thermal-conductivity polymer-based composites and provides new insights into the interpretable application of ML in materials science.

Key words: Machine learning, Residual-corrected strategy, Highly thermally conductive composites, SHAP interpretability analysis, High-throughput design