SANE: A Neural Surrogate for Accelerated Multi-Topology Analog Circuit Simulation with Spectral-Aware Training

Analog circuit simulation with numerical ODE solvers is reliable, but repeated transient analysis can be too slow for parameter sweeps, tolerance studies, and optimisation. This paper presents \textbf{SANE} (\textbf{S}pectral-\textbf{A}ware \textbf{N}eural \textbf{E}mulation), a single causal Temporal Convolutional Network (TCN) trained to emulate transient waveforms for \textbf{21 analog circuit topologies}. The loss combines time-domain $\ell_1$ error with multi-resolution STFT terms on spectral magnitude and phase. This helps the shared model capture both waveform shape and frequency content across different circuits. The dataset contains \num{176100} simulated waveform pairs spanning passive filters, diode and transistor circuits, operational amplifier stages, and switching converters. On this set, SANE reaches a mean NRMSE of 8.39\% across all topologies. After training, inference takes \SI{0.127}{\milli\second} per sample on an RTX~5090. The reference solver takes from milliseconds to tens of seconds on a single CPU core, corresponding to a mean speed-up of about ${\sim}38{,}000\times$ across the library. Because the model is differentiable, it can also be used for gradient-based parameter optimisation.