Abstract Tropical cyclones (TCs) generate pronounced sea surface cooling through multiscale air‐sea interactions, yet predicting its full evolution and identifying the governing mechanisms remain challenging. Here we develop TC‐Net, an interpretable deep learning framework that integrates a 3D convolutional encoder with a Transformer bottleneck to predict TC‐induced sea surface temperature anomalies. Trained on multi‐sensor observations from 545 Northwest Pacific TCs (1998–2022), TC‐Net produces accurate 14‐day predictions ( R 2 ≈ 0.89) and reproduces key spatial features such as enhanced right‐side cooling. To explain these predictions, we apply Integrated Gradients and derive dynamic, phase‐resolved attributions. The results reveal a robust three‐stage control sequence: (a) early cooling dominated by wind‐driven mixing and pre‐TC thermal structure (days 0–1); (b) intermediate modulation by mesoscale eddies with clear polarity‐dependent thermocline effects (days 2–8); and (c) late‐stage recovery governed by subsurface heat content (days 9–13). These attribution patterns align with the established physical understanding of TC‐ocean interactions and quantify how the dominant mechanisms evolve over the cold‐wake lifecycle. Together, the prediction and attribution results demonstrate that explainable deep learning can provide both skillful SST‐cooling predictions and physically interpretable diagnostics. TC‐Net thus offers a complementary, data‐driven approach for understanding TC‐induced ocean responses and for improving coupled forecasting and process‐based analyses.
Li et al. (Sun,) studied this question.