Multisite evaluation of physics-informed deep learning for permafrost prediction in the Qinghai-Tibet Plateau

Abstract

Prediction of the permafrost ground temperature is challenging due to its complex nonlinear process. Coupling physical models and machine learning has shown great potential in big data analysis, but its performance in permafrost prediction is still unclear. In this study, we proposed a physics-informed deep learning framework (i.e., PI-LSTM) to predict permafrost ground temperature profiles. The framework combines physical information with a long short-term memory (LSTM) network by two-stage training (pretraining and fine-tuning). The output of a Geophysical Institute Permafrost Laboratory 2 (GIPL2) model was used to pretrain the LSTM model. Borehole ground temperature measurements were used to fine-tune the pretrained LSTM model. Validation at multiple sites in various situations in the Qinghai-Tibet Plateau (QTP) shows that the accuracy and efficiency of the PI-LSTM model are dramatically higher than those of the original LSTM or GIPL2 model. Depending on the fine-tuning data sampling frequencies, the performance of the PI-LSTM model improved by an average of 27% (6–56%) and 69% (64–71%) compared to the LSTM and GIPL2 models, respectively. Even when only one year of observations was used to fine-tune the model, the RMSE of the simulated near-surface ground temperature profiles in the next decade did not substantially decline. The incorporation of physical information also contributes to the simulation efficiency. The PI-LSTM model converges faster than the LSTM model, and the training epochs are reduced by 70% (67–73%) during the fine-tuning step. This study demonstrates that integrating physical information into deep learning is promising for improving permafrost predictions.