Performance-Carbon Trade-Offs across Architectural Biases in Shear Flow Forecasting

Development of modern deep learning methods has been driven primarily by the push for improving model efficacy (accuracy metrics), leading to large-scale models that require massive computational resources and result in considerable carbon footprint across the model lifecycle. In this work, we explore how architectural biases, specifically a model's receptive field and periodicity assumption, are associated with the trade-offs between predictive performance and carbon footprint for spatio-temporal forecasting of incompressible shear flow. We study seven models differing in these properties and evaluate pointwise accuracy, physics-fidelity, and carbon cost across training and inference. We find that no single model dominates across both predictive performance and carbon cost, that a lower training cost does not straightforwardly extend to inference, and that runtime alone is an unreliable proxy for carbon cost, particularly during training. Together, these results underscore the importance of explicitly evaluating carbon costs across the full model lifecycle. We argue that model efficiency, alongside efficacy, should be a core consideration in machine learning model development and deployment.