ML-Augmented Circuit Simulation for Nonlinear Accuracy

The patent introduces machine learning models as intermediary components between the conventional circuit simulator and the final simulation results. These ML models are trained to predict nonlinear effects (such as AM/AM and AM/PM distortion, memory effects, and thermal effects) that are difficult to capture with traditional mathematical models. The ML models receive simulation inputs and outputs and augment them with predicted nonlinear characteristics, thereby improving accuracy without requiring the core simulator to become more complex.

The patent transforms the simulation approach by changing from pure mathematical modeling to a hybrid approach that incorporates machine learning parameters. The ML models are trained on measured data to learn complex nonlinear relationships, and their parameters (weights, biases, activation functions) are optimized to match real-world behavior. This parameter transformation allows the system to capture nonlinear dynamical effects that traditional fixed mathematical models cannot represent.

The patent applies preliminary action by training machine learning models offline using measured data from actual circuit operation. During the training phase, the ML models learn complex nonlinear relationships and are stored for later use. When performing actual simulations, the pre-trained models quickly predict nonlinear effects without requiring intensive real-time computation. This separates the computationally intensive learning phase from the simulation phase, reducing simulation time while maintaining accuracy.

The patent creates simplified copies of complex nonlinear behaviors through machine learning models. Instead of using full sophisticated nonlinear simulation models during operation, the system uses trained ML models that replicate the essential nonlinear characteristics (AM/AM distortion, AM/PM distortion, memory effects, thermal effects) in a computationally efficient manner. These ML copies capture the dominant nonlinear effects without requiring the full computational burden of detailed physical models.

The patent implements self-service by enabling the simulation system to automatically improve its own accuracy through machine learning. The ML models are trained on measured data and automatically learn the corrections needed to match real-world behavior. This eliminates the need for manual backfitting and trial-and-error tuning by designers. The system self-adjusts by using the trained ML models to predict and correct nonlinear effects, memory effects, and thermal effects automatically during simulation.

The patent incorporates feedback mechanisms where machine learning models are trained using measured data from actual circuit operation. The measured results provide feedback to the ML training process, allowing the models to learn from real-world performance and continuously improve simulation accuracy. This data-driven feedback loop replaces manual tuning with automated learning, reducing development time while achieving accurate predictions across multiple operating conditions.