TCAD is where device iteration stalls. A 3D FinFET drift-diffusion simulation escalates steeply with mesh complexity; multi-objective design space exploration is limited to however many full runs a team can afford in a schedule. Existing ML surrogates map scalar design parameters to scalar device metrics, which discards the physics and makes the model brittle across geometry changes. This surrogate operates on the tetrahedral mesh directly as a graph and predicts the fundamental unknowns of the drift-diffusion system -- electrostatic potential, electron and hole quasi-Fermi levels -- at every mesh node. The physics residuals are in the training objective, not bolted on after.
The consequence of operating on the mesh is size generalization: a model trained on few-fin structures applies unchanged to larger arrays, bounded only by GPU memory. That breaks the retraining cycle that makes surrogate-based DSE expensive to set up. The active-learning loop is the operational piece: per-node uncertainty from a deep ensemble identifies which candidate geometries are informationally valuable and which are redundant. Large candidate pools screen in seconds; full TCAD simulation runs only on the fraction that actually moves the Pareto front.
Device teams iterating on FinFET geometry variations have been scheduling TCAD runs like lab time -- blocks on a calendar rather than CI jobs. A surrogate that generalizes across geometry and screens in seconds moves that constraint. The bottleneck shifts from "how many simulations can we afford?" to "can we interpret the Pareto front fast enough?" That is a better problem to have. Accepted at ICCAD 2026.