Frequency-Band-Aware Physics-Informed Generative Adversarial Network for EMI Prediction and Adaptive Suppression in SiC Power Converters

Silicon carbide (SiC) power converters offer superior switching performance but generate severe broadband electromagnetic interference (EMI) that challenges regulatory compliance. Existing prediction methods face a fundamental trade-off between physical fidelity and computational efficiency, while conventional suppression strategies lack adaptability to varying operating conditions. This paper proposes a frequency-band-aware physics-informed generative adversarial network (FBA-PIGAN) that integrates electromagnetic domain knowledge into data-driven generative modeling for joint EMI prediction and adaptive suppression in SiC power converters. The framework employs a Wasserstein GAN with gradient penalty as the adversarial backbone and introduces feature-wise linear modulation (FiLM) to inject converter operating parameters into the generator through learned affine transformations. A hierarchical physics-informed loss function enforces three frequency-dependent constraints, namely, harmonic structure consistency, parasitic resonance characterization, and high-frequency envelope regularization, coordinated by a curriculum-based weight-scheduling strategy. An end-to-end differentiable suppression module maps predicted spectra to optimal passive filter parameters through an analytically embedded transfer function. Experimental validation on a 10 kW SiC inverter platform with 5120 measured spectra across 32 operating conditions demonstrates that FBA-PIGAN achieves a mean spectral error of 2.1 dB, 93.8% peak frequency accuracy, and a physical consistency score of 0.93, improving prediction accuracy by 56% over conventional conditional GANs while maintaining sub-millisecond inference latency. The integrated suppression pipeline attains 19.2 dB average attenuation with 98.5% CISPR 25 compliance, and the framework generalizes to unseen operating conditions with only 19% performance degradation, compared with 56% for data-driven baselines.