This study , led by Cheng Zhang, improves machine-learned models for simulating ocean mesoscale eddies by addressing artifacts near coastlines caused by out-of-sample predictions using convolutional neural networks (CNN). Comparing two different strategies for treating Boundary Conditions (BCs) in CNN models - zero and replicate padding - they show that replicate padding significantly reduces these artifacts, enhancing the accuracy and stability of ocean modeling near complex coastal regions.