Surrogate Assisted Compact Model Extraction for Commercial CMOS Process Operating at Cryogenic Temperatures

Cryogenic circuit design requires compact transistor models that reproduce measured low-temperature behavior and remain compatible with circuit simulators. We find BSIM4 parameter values that fit experimentally measured I--V curves using a neural surrogate of NGSpice. The surrogate maps compact-model parameters and device geometry to complete output and transfer characteristics. Extraction then proceeds as a bounded inverse problem: a bank of NGSpice simulations provides initial candidates, finite differences of the surrogate guide local least-squares refinement, and the resulting cards are re-simulated in NGSpice before selection. We compare searches over seven effective card parameters and a sensitivity-screened set of 43 parameters across 18 measured geometries. Relative to the re-simulated cards from the previous extraction, the two searches reduce in-sample mean relative root-mean-square error (RRMS) by 17\% and 26\%, respectively. The larger search has the lower overall error, although the seven-parameter fit remains better in some bias regions and individual curves regress. Code is available at \url{https://anonymous.4open.science/r/cryo-compact-model-F01D/}.