Predicting Soil Temperature and Quantifying Thermal Effects of Straw Mulching via a Data-Physics Dual-Driven Neural Network (Hy-PINN)

Soil temperature is a critical meteorological factor influencing crop growth. Given that the dynamics of soil temperature under various straw mulching practices remain unclear, a two-year field experiment (2024–2025) was conducted. Four treatments were established—no mulching (CK), chopped straw mulching (CS), strip tillage (SS), and whole straw mulching (WS). Results indicated that straw mulching significantly altered the soil thermal environment; the cooling intensity followed the order of WS > SS > CS (averaging 0.58–1.72 °C) and was primarily concentrated during the seedling stage (1.09–2.82 °C). While full-scale mulching may induce early cold stress, SS mitigated this effect, providing a more balanced thermal environment. To address the reliance of traditional physical models on soil boundary conditions and parameters, we proposed a data-physics dual-driven hybrid physics-informed neural network (Hy-PINN). This model simultaneously predicts soil temperatures at 0 cm and 10 cm depths and dynamically inverts the apparent thermal diffusivity (κ). Through strict cross-year evaluation (training on 2024, independent testing on 2025), the Hy-PINN achieved high-precision predictions across all scenarios (R2 ≥ 0.89; RMSE ≤0.95 °C at 0 cm and ≤1.10 °C at 10 cm). The Hy-PINN model serves as a robust complementary tool to field monitoring, providing a scientific basis for assessing the thermal suitability of conservation tillage patterns.