Do machine learning climate models work in changing climate dynamics?

Our baseline runs followed the ClimateSet single emulator specifications (Kaltenborn et al., 2023): Training Process: Each emulator is trained on data from a single climate model, predicting outputs for an entire sequence of monthly data for each year. Pre-Processing: The data has been pre-processed by ClimateSet to have a spatial resolution of approximately 250 km (144 x 96 longitude-latitude cells) and a temporal resolution of monthly data. The time series is divided into 1-year chunks, resulting in data with a shape of scenarios, years * months, variables, longitude, latitude . Input and Output Shapes: The input data has the shape batch, sequence length, num vars, lon, lat, where the sequence length is 12 (monthly data). The output has the shape batch, sequence length, 2, lon, lat, where the '2' corresponds to temperature (T AS) and precipitation (PR). Training Parameters: The models are trained for 50 epochs with an initial learning rate of 2e-4, using an exponential decay scheduler. For the non-frozen ClimaX models, training begins with a 5-epoch warm-up phase at 1e-8, followed by training at 5e-4. Loss: The latitude-longitude weighted mean squared error (LLMSE) as implemented in (Nguyen et al., 2023) is used.