Machine Learning Electric Drive Unit Simulation

The patent applies preliminary action by pre-training machine learning models offline using comprehensive multi-physics simulation data. The neural networks are trained beforehand on datasets generated from detailed electromagnetic, thermal, and mechanical simulations, storing learned patterns and relationships. During actual design optimization, the pre-trained models provide rapid predictions without requiring real-time execution of computationally expensive multi-physics simulations, thus achieving both high accuracy and fast performance evaluation.

The patent implements feedback mechanisms where the machine learning models' predictions are continuously validated and refined. The system uses active learning approaches where uncertain predictions trigger additional training iterations, and prediction results feed back into model refinement processes. This feedback loop allows the models to adapt to new parameter ranges and maintain high accuracy even when input parameters differ significantly from the original training dataset, while preserving the computational speed advantages of ML-based prediction.

The patent replaces the traditional mechanical/computational multi-physics simulation system with machine learning-based prediction systems. Instead of executing complex coupled differential equations during each optimization iteration, the system uses pre-trained neural networks that have learned the underlying physical relationships. This substitution maintains design accuracy by preserving the physics-informed nature of the models while dramatically improving optimization efficiency, enabling rapid evaluation of multiple design configurations and facilitating gradient-based optimization methods.