Bridging physical constraints and deep generative models via physics-aware normalizing flows

Generative artificial intelligence suffers from critical limitations including hallucination and violation of real-world constraints. To address this, we present a physics-aware generative modeling framework combining adversarial learning with physical forward models and Normalizing Flows. Systematic evaluation across simple to complex electric circuit models and four Normalizing Flow architectures demonstrates broad applicability. Training stabilization is achieved through discriminator ensembles, instance noise scheduling, and penalty terms. Downstream supervised learning tasks validate the framework’s ability to synthesize labeled training data from unlabeled impedance measurements, achieving competitive performance across several circuit elements compared to original datasets. • Physics-Aware Normalizing Flows integrate adversarial learning with physical models. • Framework guarantees physically plausible data via embedded simulator. • Labeled training data is synthesized from unlabeled impedance measurements. • Four Normalizing Flow architectures are evaluated across simple to complex equivalent circuits. • Advanced training stabilization enables convergence for complex physical models.