We have developed a purely convolutional neural network to increase the spatial resolution of climate models, necessary to provide climate change scenarios at the local scale. Training is conducted using PyTorch. Results have been assessed against HCLIM, a high-resolution regional climate model, showing comparable results. This confirms that the signal of a coarse-resolution global climate model can be used with these methods to provide regional climate information. The method is evaluated against CERRA and against HCLIM, an even higher-resolution regional dynamic climate model (3 km) divided into three European subregions. It has now been applied using HCLIM as ground truth, for application to future projections. First results on its extrapolability to future climates were presented at the European Geosciences Union annual meeting. We have also contributed our convolutional neural network to the recently submitted ML-CORDEX intercomparison paper, which benchmarks machine-learning downscaling methods across multiple research groups worldwide. Building on this foundation, we developed CDSI, a Neural-LAM-based stochastic interpolant method for downscaling, and benchmarked it against two other machine-learning approaches, CorrDiff and UNet, as well as against HCLIM, over the EURO-CORDEX domain. CDSI reproduces extreme precipitation and temperature events comparably to, and in several metrics better than, the other machine-learning methods and the dynamical regional model itself, better preserving large-scale spatial structure while retaining the fine-scale detail needed to capture distribution tails. These results were presented at the AI4PEX General Assembly in Brest and are being consolidated into publications, one submitted and another in preparation. Given this performance, we intend to continue developing CDSI as our primary emulation method. The high-resolution datasets generated with this method are used to assess how different global warming levels, rapid climate system changes, and tipping points impact regional climate, focusing on Europe and polar regions, and assessing changes in mean state, variability, and extremes. For this phase, computing requested on Arrhenius will support two related lines of work: We will train our downscaling models on CMIP6 climate model data and apply them to downscale idealized simulations representing +2K and +4K global warming levels over Europe, produced within OptimESM. Training across multiple driving models and warming levels tests the method's robustness and transferability beyond its original historical/reanalysis-driven setting, and characterises how regional extreme statistics shift with warming. We are extending CDSI, our multi-variable, multi-ensemble-member regional climate model emulator developed with Linköping University (based on their Neural-LAM graph-based neural weather prediction framework), to emulate HCLIM output directly. The emulator produces 13 output variables, well beyond the precipitation and temperature fields of our initial work, allowing ensembles of high-resolution, HCLIM-consistent fields to be generated at a fraction of the dynamical model's computational cost. Results generated on Arrhenius will provide high-resolution information to understand regional threats from different warming levels, part of the EU Horizon projects OptimESM and AI4PEX. Development and initial testing were carried out on Berzelius, with subsequent production on Alvis; with Alvis's retirement, we are now applying for computing time on Arrhenius to continue these projects.