Evaluating Pre-Loss Normalization Across Multiple Loss Functions in Deep Learning for Multivariate Fluid Flow Prediction

Multi-scale flow fields can destabilize the training of data-driven models by weighting loss contributions unevenly. This study proposes a loss-independent pre-loss normalization module, used only during training, that independently scales predictions and targets prior to loss computation. Three normalization schemes are tested with four loss functions using a transformer model for laminar developing pipe flow. On the test dataset and two additional flow cases, Pre-Loss Normalization reduces the mean absolute error by 27–81% and improves the R2 value of 0.186–0.630 over unnormalized training. Overall, this approach mitigates multi-scale effects and improves accuracy and robustness.