Aiming to address the challenge of high-fidelity modeling and rapid parameter optimization for machine - crop - soil coupling systems under small-sample conditions, taking the cotton stalk pulling device as an example, a cross-scale multi-objective optimization framework that integrated the discrete element method (DEM) with physics-informed neural networks (PINN) was proposed. The framework consisted of DEM-based microscale modeling, mechanical relation extraction, PINN physical-constraint embedding, and NSGA -II optimization. Firstly, a composite "rubber roller - cotton stalk - soil" DEM model was established, from which the stick - slip - separation contact mechanism was extracted. The geometric indentation and equivalent contact stiffness were merged into an identifiable parameter and incorporated into the PINN as a primary physical constraint to ensure physical consistency. Subsequently, with stalk breakage rate and miss-pulling rate as dual objectives, NSGA -II was employed to compute the Pareto front within engineering feasibility bounds and supported by sensitivity analysis, and the knee point was selected as the optimal operating condition. Results showed that the proposed framework reduced the computational cost per case from approximately 8 h by using DEM to 2 s by using PINN forward inference, achieving a validation R2=0.95 and improving prediction accuracy by 20% ~ 30% compared with conventional neural-network surrogates. The optimal operating condition corresponded to a breakage rate of 8% and a miss-pulling rate of 7%, with a field-test deviation of less than 2%. Overall, the proposed DEM - PINN cross-scale framework maintained physical consistency while significantly reducing optimization cost, demonstrating strong generalizability and transferability for soil - plant - machine coupling systems.