Physics-based hydrodynamic models are essential for accurate flood prediction but are computationally expensive, limiting their applicability for real-time forecasting and probabilistic analyses. Conversely, pure machine learning (ML) models offer both computational efficiency and accuracy but often lack interpretability. To address this gap, we propose SGUnet, a physics-informed ML model and a hybrid theory-guided data science approach, for rapid, high-resolution flood mapping. It utilizes a neural network with U-Net architecture and integrates subgrid-based coarse-grid hydrodynamic model predictions as initial estimates, upskilling them to achieve fine-grid model accuracy. Unlike traditional hydrodynamic models, the subgrid method embeds fine-scale topographic details within coarse-grid cells, enhancing both computational efficiency and predictive accuracy. SGUnet processes flood depth raster patches (512 × 512 pixels) and corresponding digital elevation models as inputs. It functions as a deep learning-based corrector, refining flood predictions from numerical simulators. Trained through supervised learning, SGUnet learns to correct deviations in coarse-grid predictions using fine-grid model outputs as target values. The model is evaluated across three large Australian watersheds—Wollombi, Chowilla, and Burnett River—using HEC-RAS flood simulations with subgrid formulation. SGUnet reduces root mean squared error by a factor of 4.5–5.3 compared to coarse-grid models, achieves a critical success index exceeding 0.9 for flood extent mapping, and delivers a 50x speed-up over fine-grid hydrodynamic models. Furthermore, SGUnet outperforms a state-of-the-art ML-based upskilling model in depth and extent predictions. By effectively correcting flood artifacts from coarse-grid models, SGUnet achieves near fine-grid accuracy with significantly reduced computational cost, demonstrating its potential for real-time flood risk assessment.

Machine learning (ML) applications in Earth and environmental sciences (EES) have gained incredible momentum in recent years. However, these ML applications have largely evolved in ‘isolation’ from the mechanistic, process-based modelling (PBM) paradigms, which have historically been the cornerstone of scientific discovery and policy support. In this perspective, we assert that the cultural barriers between the ML and PBM communities limit the potential of ML, and even its ‘hybridization’ with PBM, for EES applications. Fundamental, but often ignored, differences between ML and PBM are discussed as well as their strengths and weaknesses in light of three overarching modelling objectives in EES, (1) nowcasting and prediction, (2) scenario analysis, and (3) diagnostic learning. The paper ponders over a ‘coevolutionary’ approach to model building, shifting away from a borrowing to a co-creation culture, to develop a generation of models that leverage the unique strengths of ML such as scalability to big data and high-dimensional mapping, while remaining faithful to process-based knowledge base and principles of model explainability and interpretability, and therefore, falsifiability.