DeepCUN: MATLAB code for Baltic Sea surface-current emulation, explainability, and autoregressive forecasting This Zenodo release provides the MATLAB workflow used to preprocess data, train and test the DeepCUN deep convolutional U-Net emulator for Baltic Sea daily mean sea-surface currents (SSC), compute post hoc explainability diagnostics (LRP and DJE), and run autoregressive multi-day forecasting experiments. The code supports the analyses and figures reported in the accompanying manuscript. What is included The workflow is organized as sequential scripts (Steps 0–4). Plotting scripts are provided to reproduce the final figures from the XAI (Step 3) and forecasting (Step 4) outputs. Core pipeline Step_0_Data_preparation.mPreprocessing and dataset assembly: domain cropping, land–sea masking, daily aggregation, and normalization/scaling of atmospheric variables to the SSC range as described in the manuscript. Step_1_DeepCUN_traninig_final.mModel training for DeepCUN. Configures training, runs optimization, and saves model weights/checkpoints. This release also includes the final trained model weights as DeepCUN_2vars.mat, enabling users to skip training and reproduce testing/XAI/forecasting directly. Step_2_DeepCUN_testing.mOne-step testing and evaluation on the independent test year. Computes performance metrics (e.g., MAE, CC, ED) and saves evaluation outputs used in the manuscript. Explainability (XAI) Step_3a_DeepCUN_XAI_LRP.mLayer-wise Relevance Propagation (LRP) computation and saving of relevance maps/statistics. plot_LRP_finalFIG.m (run after Step 3a)Generates the final LRP figure(s) used in the manuscript from the saved Step 3a outputs. Step_3b_DeepCUN_XAI_DJE.mDiagonal Jacobian Elasticity (DJE) computation (Hutchinson/Rademacher probing) and saving of elasticity maps/statistics. plot_DJE_finalFIG.m (run after Step 3b)Generates the final DJE figure(s) used in the manuscript from the saved Step 3b outputs. Autoregressive forecasting Step_4_forecast_generator.mGenerates autoregressive rollouts (multi-day forecasts) by recycling predicted SSC as input and prescribing wind forcing at the corresponding lead time. Step_4_forecast_windows_analyser.mComputes lead-time skill curves and spatial skill diagnostics across selected horizons/windows. plot_pred_leadtime_analyser.m (run after Step 4 analyses)Produces the final lead-time performance plot(s) (e.g., MAE/CC vs lead time) used in the manuscript. Data requirements (downloadable via included scripts) This release includes all input data in the CMEMS/ and ERA5/ folders required to run the full DeepCUN. For completeness, helper scripts are also included to re-download the original datasets from the official providers if needed (Copernicus Marine Service and the Copernicus Climate Data Store).. Recommended execution order Step 0 → prepare processed inputs and masks Step 1 → train DeepCUN and save weights Step 2 → test on the independent year and save metrics/fields Step 3a, then plot_LRP_finalFIG.m Step 3b, then plot_DJE_finalFIG.m Step 4 generator + analyser, then plot_pred_leadtime_analyser.m Software environment MATLAB with required toolboxes (typically including Deep Learning Toolbox). GPU acceleration is supported if available.