Work

Neural SPICE Surrogate Modeling and Sensitivity Analysis for Semiconductor MOSFETs: A Deep Learning Framework for Accelerated VLSI Simulation

Year: 2026
Type: preprint
Abstract: Modern Very Large Scale Integration (VLSI) electronic design automation (EDA) workflows rely heavily on circuit simulation engines (SPICE) to resolve large systems of non-linear differential-algebraic equations. As semiconductor dimensions scale, computing drain current characteristics iteratively via traditional numerical techniques introduces severe computational latency, particularly across extensive Monte Carlo process-variation sweeps. In this work, we propose a deep learning surrogate modeling framework to approximate device-level non-linear transport physics. The surrogate establishes a direct forward mapping F[๐‘‰๐‘”๐‘ , ๐‘‰ds, ๐‘‰๐‘กโ„Ž, ๐œ‡0, ๐œ†]--> Id across all operating regimes (cutoff, linear, and saturation) utilizing a high-capacity Multi-Layer Perceptron (MLP) architecture equipped with continuous Sigmoid Linear Unit (SiLU) activations. Trained on synthetically generated physical parameter profiles, the model demonstrates rapid convergence, driving means squared loss from an initial 0.122 down to 0.0078 with a coefficient of determination R^2 > 0.98 on unobserved test partitions. Furthermore, a normalized sensitivity analysis identifies gate bias ๐‘‰๐‘”๐‘  and threshold voltage ๐‘‰๐‘กโ„Ž as primary driving variables, fully consistent with metal-oxide-semiconductor field-effect transistor (MOSFET) electrostatics. The proposed surrogate achieves evaluation latencies under 2 milliseconds, providing a scalable path toward accelerated analog, digital, and post-silicon statistical simulations.
Language: English
Open Access status: green
OpenAlex ID: W7236846141
Cites: 5
Cited by:
Related to: 10