Surrogate-Assisted Compact-Model Extraction for SKY130 CMOS at 77 K

Aaron Wang, Olivia Gzamouranis, Vaibhav Lohia, Hershel Fei Xun Weiner, Olivia Seidel, Richard Cavanaugh, Giuseppe Di Guglielmo, Nhan Tran, Davide BragaarXiv 2026 · arXiv

TL;DR

Researchers have developed a machine-learning-assisted pipeline to calibrate transistor models for cryogenic temperatures (77 K). By using a neural network as a "surrogate" for the NGSpice simulator, they achieved up to a 26% improvement in modeling accuracy for SkyWater 130 nm CMOS devices compared to existing industry-standard models.

What Problem Does This Paper Address?

As we push toward large-scale quantum computing and high-energy physics experiments, we need electronics that can operate at cryogenic temperatures (like 77 K). Standard "compact models" (the mathematical descriptions of transistors used in circuit simulators) are typically optimized for room temperature. When cooled, a transistor's behavior—such as its threshold voltage and electron mobility—shifts significantly.

Recalibrating these models for the "cold" regime is notoriously difficult. It requires fitting dozens of parameters to experimental data, a process that usually involves thousands of slow calls to a circuit simulator. Furthermore, the parameter space is "ill-posed," meaning many different combinations of settings can produce similar-looking results, making it hard to find the "true" physical model.

Methods

The authors propose a Surrogate-Assisted Extraction pipeline that replaces slow simulator calls with a fast, differentiable neural network.

  1. Build a Simulator Bank: They first generate 10,000 random parameter sets using Latin Hypercube sampling and simulate them in NGSpice to create a local "map" of the parameter space.
  2. Train a Forward Surrogate: A multilayer perceptron (MLP) is trained on this bank. This surrogate learns to predict the entire I-V (current-voltage) characteristic of a transistor based on its geometry and model parameters.
  3. Local Refinement: Instead of calling the simulator during optimization, they use the surrogate. Because the neural network is differentiable, they can use its Jacobian to perform a fast, bounded least-squares search to find the best fit for measured 77 K data.
  4. Final Confirmation: To ensure accuracy, the best candidates proposed by the surrogate are re-simulated in the actual NGSpice engine for final validation.
Figure 1: Extraction pipeline. NGSpice generates the simulation bank, the frozen forward surrogate is used for local refinement, and NGSpice re-simulates the selected candidates for comparison with measurement.

Datasets

The study uses experimental measurements of SkyWater 130 nm (SKY130) CMOS transistors at 77 K. The dataset includes:

  • 18 unique geometries: 8 nMOS and 10 pMOS devices.
  • 198 total I-V curves: Covering both output and transfer characteristics.
  • Baseline: They compare their results against the previously released 77 K model cards from Beall et al. (2024).

Key Insights and Ideas

  • The Bank vs. The Refinement: Interestingly, the authors found that the broad initial search in the "simulator bank" provided most of the accuracy gains. The surrogate-guided refinement added a smaller, but still significant, fine-tuning step.
  • Parameter Sensitivity: They compared a standard 7-parameter search (focusing on known temperature-sensitive variables) against an expanded 43-parameter search. While the 43-parameter model was more accurate overall, the 7-parameter model actually performed better in "high-bias" regions (high voltage/current), suggesting that more parameters can sometimes lead to overfitting in specific operating regimes.
  • Differentiable Physics: By using a neural surrogate, the team could calculate "gradients" (directions to move parameters to improve the fit) without the simulator providing them, effectively turning a "black-box" simulator into an optimizable function.

Key Results

The researchers measured success using the Mean Relative Root-Mean-Square Error (RRMS) across all devices.

  • Overall Improvement: The 7-parameter search reduced error by 17%, while the 43-parameter search reduced it by 26% relative to the baseline.
  • nMOS vs. pMOS: The improvements were particularly stark for nMOS devices, where the error dropped from 0.120 to 0.070.
Figure 2: Mean device RRMS from NGSpice re-simulations of the released, seven-parameter, and 43- parameter cards. “High bias” averages, within each device, the output curves at fixed |VGS| ≥1.48 V and the transfer curves at fixed |VDS| ≥1.48 V.

As seen in the figure below, the 43-parameter model (red line) tracks the measured saturation current much more closely than the original released model (dashed blue line).

Figure 3: Output characteristic of the nMOS L = 1 µm, W = 3 µm device at VGS = 1.85 V. Left: full VDS sweep. Right: VDS ≤0.37 V. Lines are NGSpice re-simulations; markers are measurements.

Limitations and Future Work

  • In-Sample Evaluation: The models were evaluated on the same 18 geometries used for sensitivity screening. It remains to be seen how well this approach generalizes to entirely unseen transistor sizes.
  • DC Only: The current work focuses on DC (steady-state) behavior. Future work is needed to address dynamic response, noise, and variability, which are critical for high-speed cryogenic circuits.
  • Regressions: Despite the overall improvement, about 20% of individual curves actually showed higher error than the baseline (see Figure 4 in the paper). This highlights the "tug-of-war" in global fitting where improving one region of operation can inadvertently degrade another.

Summary updated

Abstract

Cryogenic circuit design requires compact transistor models that reproduce measured low-temperature behavior and run in standard circuit simulators. We fit BSIM4 parameters to measured I-V curves of SkyWater 130 nm (SKY130) transistors at 77 K using a neural surrogate of NGSpice. The surrogate maps compact-model parameters and device geometry to the complete set of output and transfer characteristics. Extraction is a bounded least-squares search: a bank of NGSpice simulations supplies initial candidates, finite differences of the surrogate supply the local Jacobian, and NGSpice re-simulates the resulting cards before selection. On 18 measured geometries, we compare a search over seven card parameters with a search over 43 parameters selected by NGSpice sensitivity. Relative to NGSpice re-simulations of the previously released cards, the two searches reduce the in-sample mean relative root-mean-square error (RRMS) by 17% and 26%, respectively. The 43-parameter search has the lower overall error, but the seven-parameter search has the lower mean on high-bias curves, and some individual curves have higher error than with the released cards under both searches.
BibTeX
@misc{wang2026surrogate,
  author = {Aaron Wang and Olivia Gzamouranis and Vaibhav Lohia and Hershel Fei Xun Weiner and Olivia Seidel and Richard Cavanaugh and Giuseppe Di Guglielmo and Nhan Tran and Davide Braga},
  title = {{Surrogate-Assisted Compact-Model Extraction for SKY130 CMOS at 77 K}},
  year = {2026},
  eprint = {2611.21594},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21594},
}
Figure