Deep learning models, particularly Long Short-Term Memory (LSTM) networks, achieve state-of-the-art overall performance in regional rainfall–runoff modeling. However, under unprecedented extreme meteorological forcings, purely data-driven LSTMs can suffer from artificial output saturation—a critical structural limitation that remains largely hidden during standard historical evaluations. In this study, we utilize a regional, non-stationary weather generator to produce synthetic extreme precipitation events, stress-testing a standard LSTM alongside mass-conserving conceptual HBV and hybrid models across a subset of catchments from the CAMELS-DE dataset. To mitigate the LSTM’s saturation problem, we introduce a targeted data augmentation strategy. By driving the physics-constrained models with synthetic extreme forcings from the weather generator, we create physically plausible discharge targets to enrich the LSTM’s training data.</p> <p>Our results demonstrate that adding even a small fraction of synthetic extreme events successfully mitigates artificial output saturation by shifting the LSTM’s predictive limit upward, even exceeding that of an LSTM trained on the entire CAMELS-DE dataset. The augmentation improves peak-flow simulation performance without compromising overall simulation accuracy. We validate this approach using the catastrophic out-of-sample July 2021 flood event across multiple affected gauges, demonstrating that regional data augmentation can significantly reduce peak flow underestimation. However, while augmentation mitigates the practical impact of the LSTM's architectural ceiling, it does not eliminate it. Notably, the evaluated hybrid model natively avoids saturation, yielding the best overall extreme-flow performance. Ultimately, for operational flood forecasting and design flood estimation, we emphasize the necessity of quantifying the LSTM's predictive limit and monitoring model outputs for artificial saturation. To prevent the potentially catastrophic underestimation of unprecedented extremes, we strongly recommend deploying mass-conserving or hybrid architectures in parallel with standard LSTMs as structural safeguards.