Long horizon paleoclimate forecasting using ice core data

Paleoclimate reconstruction is crucial in understanding long-term climate variability. Ice core records are valuable proxies for reconstructing past climate changes and provide an enhanced comprehension of long-term climate trends. However, forecasting future climate variables using ice core data remains challenging due to non-linearity, missing values, and uncertainty of paleoclimatic signals. This study proposes a hybrid AI and Physics-Informed deep learning model to improve paleoclimate reconstruction from oxygen-18 isotopes (δ18O) in ice core data. The study employs the Dome Fuji-338KYr Wet Extraction CO2 dataset, Empirical Mode Decomposition and Wavelet Transform are used for feature extraction, isolating critical climate signals. The extracted features are input into three parallel models, PINN, TFT, and BNN, which predict δ18O individually. The predictions are then combined through performance-based weighted averaging. Finally, the Dansgaard relationship is used to reconstruct the temperature of the past. The proposed models exhibit better performance in paleoclimate forecasting than other methods, with lower error rates of 0.0995 RMSE and 0.0099 MSE. The weighted average approach combines model predictions, ensuring accurate forecasts and reliable temperature reconstruction over 10 kyr.