A parallel evaluation framework integrating a modified SWAP model and AutoML for simulating soil salinity under mulched drip irrigation

Accurate characterization of soil salinity dynamics in arid mulched drip-irrigated cotton fields is essential for water-salt regulation and salinization risk management. However, this remains challenging because plastic film mulching reshapes surface water-salt fluxes and enhances field-scale spatial heterogeneity. Although physical models and machine learning approaches have been widely used for salinity simulation, their applicability remains constrained by simplified film-boundary representations, parameter spatial heterogeneity, manual model selection, and limited interpretability. To address these issues, this study developed a parallel evaluation framework integrating a modified SWAP model and H2O AutoML for soil salinity simulation. For the physical model, a film coverage factor was introduced to reformulate the upper-boundary flux mechanism, thereby establishing a SWAP model adapted to mulched cotton fields. For the data-driven model, H2O AutoML was used to automate model structures optimization and hyperparameters tuning, and variable-importance analysis was performed to identify the dominant factors controlling salt transport under mulched drip irrigation. Model accuracy was evaluated under on three scenarios: temporal extrapolation, spatial generalization, and vertical reconstruction. The modified SWAP model reduced the RMSE by 0.14 g·kg −1 compared with the standard SWAP model, indicating that the film-induced boundary flux modification improved the physical simulation of soil salinity. In the temporal extrapolation scenario, the modified SWAP model outperformed AutoML by better representing the time-lag effect of deep salt transport. In contrast, in the spatial generalization scenario, H2O AutoML reduced the NRMSE across the 0–80 cm soil layer by 46.79% relative to the modified SWAP model, demonstrating its advantage in handling plot-scale spatial heterogeneity. In the vertical reconstruction scenario, the recursive strategy constructed a vertical information chain that alleviate the attenuation of surface information during downward transmission. Compared with the direct mapping method that neglects interlayer correlations, this strategy reduced the RMSE of deep-layer salinity at 60–80 cm by 21.69% and controlled the profile-wide NRMSE to within 15%, achieving high-precision reconstruction of the salinity profile. These findings highlight the value of process-based models for long-term salinity extrapolation and risk assessment, and the advantage of AutoML in capturing spatial heterogeneity and reconstructing salinity profiles. Their complementarity provides practical support for hybrid physics-data modeling and precision water-salt management in arid agricultural systems.