Research on Thermophysical Property Prediction of High-Temperature Insulating Materials Based on Neural Network

Three models - FNN (Feedforward Neural Network), BPNN (Back Propagation Neural Network), and PINN (Physics-Informed Neural Network) - were developed to predict the thermal properties of nano-aerogel thermal insulation materials based on experimental data collected under operating temperatures ranging from 50℃ to 450℃. The accuracy and robustness of the models were verified using four statistical indicators: correlation coefficient (R), mean absolute error (MAE), root-mean-square error (RMAE), and relative error (RE). Both the BPNN and PINN models accurately predicted the thermal properties of nanoaerogel thermal insulation materials. However, the FNN model performed poorly for nanoaerogel. all three models exhibit different levels of robustness, with BPNN being the most robust, followed by PINN, and FNN showing diminished robustness. According to the analysis of prediction accuracy and error, the BPNN model demonstrates superior accuracy and robustness compared to the other two models in nanoaerogel thermal insulation materials. The PINN model exhibits slightly lower accuracy than the BPNN for nano-aerogel and exhibits certain limitations in terms of robustness, but it still outperforms the FNN model in nanoaerogel thermal insulation materials. As a result, for the prediction of thermal properties in nanoaerogel insulation materials, the BPNN and PINN models are recommended due to their better overall performance in accuracy and robustness.