• Reframes systematic build-up/wash-off biases as information-rich “memory” that can be learned and used. • Proposes a lightweight, model-agnostic machine-learning error layer that preserves mass-conserving physics. • Demonstrates in an urban catchment that lagged-error inputs yield large, robust skill gains for TSS, TN and TP. • Enables operational stormwater models to self-update in real time, maintaining reliable predictions under non-stationary conditions. Process-based build-up/wash-off models underpin urban stormwater quality management, yet structural simplifications and environmental non-stationarity often induce systematic biases that erode decision reliability. This "Making Waves" article explores a pivotal question: Can biases, long treated as model “defects”, be reframed as information assets that strengthen decision-grade predictive credibility? This paper proposes a model-agnostic machine learning error correction strategy. The core idea is to recast model-observation errors not as random noise but as structured, information-rich signals carrying “memory bias”. By learning the dynamic patterns within recent errors, this correction layer enables real-time, non-intrusive adjustments to the physical model's output. In the study catchment, it raised pollutant NSE compared baseline by roughly 20% on average with hydrologic inputs and 66% with lagged errors. Testing across two catchments demonstrates that the proposed framework achieves superior accuracy and stability compared with traditional and pure machine-learning baselines. This approach evolves traditional models from static tools requiring periodic recalibration into responsive systems capable of dynamic output correction. By making systematic biases 'learnable' and 'usable', the model can effectively respond to environmental fluctuations, ensuring its long-term validity and reliability in a constantly changing environment.
Liu et al. (Sun,) studied this question.