betaFeedback

WorksPhysics and Astronomy

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

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.

Introduction

Cryogenic CMOS places readout and control electronics close to cold detectors and qubits, which reduces the number and length of interconnects to room-temperature electronics [6, 15]. Circuit design in this setting requires compact transistor models evaluated at the operating temperature. Cooling changes the threshold voltage, mobility, subthreshold conduction, series resistance, and high-field behavior, so a room-temperature model must be recalibrated before use [2, 3, 10, 13].

A BSIM4 model card defines the drain current as a function of terminal voltages, device geometry, and compact-model parameters through equations implemented in a circuit simulator [7]. Changing only the parameter values keeps the simulator interface unchanged. Fitting those values to cryogenic measurements has two difficulties: each trial requires a simulator call, and different parameter vectors can produce similar DC curves, so the fitted vector is not unique.

We use a learned forward surrogate of NGSpice to evaluate candidate cards without simulator calls during local refinement. The fitting target is always the measured curves, and every reported score comes from NGSpice. We start from the dataset and 77 K SKY130 model cards released by Beall et al. [1], whose staged extraction fits seven temperature-sensitive multipliers and gives 18 BSIM4 models for the measured geometries. We refine those cards in two search spaces: the same seven parameters, and 43 parameters selected by NGSpice sensitivity. We measure the change in RRMS against the measured curves, how much of that change comes from surrogate-guided refinement and how much from selection from the simulator bank, and the bias regions where the fitted cards still disagree with measurement.

Related work

Cryogenic MOS modeling usually starts from an established compact model and adjusts it to low-temperature measurements. Uniform temperature scaling does not reproduce the measurements, particularly in the subthreshold region [4, 5], so published extractions change selected compact-model parameters using measurements from each technology and geometry [2, 3, 10, 13]. The SKY130 study used here keeps the room-temperature parameter dependencies through seven nominal multipliers and releases SPICE-compatible 77 K cards with its measurements [1]. Machine-learning approaches have predicted compact-model parameters directly from I–V data, including BSIM-CMG extraction for FinFET devices [8, 14]. Because this inverse mapping can assign different parameter vectors to similar curves, we instead learn the forward simulator response and perform the inverse search through the learned model [9, 12].

Experimental setup

Measurements and simulator. The released model tables contain eight nMOS and ten pMOS geometries [1, 11]. Each device has five output characteristics (IDI_D–VDSV_{DS} at fixed VGSV_{GS}) and six transfer characteristics (IDI_D–VGSV_{GS} at fixed VDSV_{DS}), for 198 measured curves in total. We simulate all cards with NGSpice 41 [16]. The baseline is a fresh NGSpice simulation of the released cards, so the baseline and both fitted card sets use the same simulator and scoring code.

Parameter sets. The seven-parameter search varies the card values of VTH0, U0, NFACTOR, VSAT, DELTA, RDSW, and ETA0 near the released 77 K card [1, 7]. The prior extraction fitted nominal multipliers on these parameters; we vary their card values directly. For the expanded search, we start from 115 continuous BSIM4 parameters of the DC-current model, perturb each one in NGSpice, and keep those whose normalized response exceeds a fixed threshold. The result is a common 43-parameter layout with 38–40 non-degenerate dimensions per device. Appendix § 11 gives the screening rule and the parameter list.

Metric. We score each curve with the relative root-mean-square error (RRMS). For measured current IkI_k and simulated current I^k\hat{I}_k on curve kk with NkN_k points,

RRMSk=Nk−1∑i(I^k,i−Ik,i)2Nk−1∑i∣Ik,i∣.(1)\mathrm{RRMS}_k = \frac{\sqrt{N_k^{-1}\sum_i(\hat{I}_{k,i}-I_{k,i})^2}}{N_k^{-1}\sum_i\lvert I_{k,i}\rvert}. \tag*{(1)}

The denominator normalizes curves with different current scales. Curve scores are averaged within each device, and device scores are averaged over the 18 devices. The released scoring code excludes 23 near-off output curves. We fix the remaining 175 curves before fitting and use the same set for every method, so a candidate cannot change its score by changing which curves are included. Appendix § 10 gives the trimming and inclusion rules.

Surrogate-assisted extraction

For a device geometry gg, NGSpice defines a forward map I=f(θ,g)I=f(\theta,g) from a BSIM4 parameter vector θ\theta to the simulated currents. Extraction searches for a θ\theta whose simulated currents match the measurement. Repeated NGSpice evaluation of ff is slow, and a learned inverse map is ill-posed because different parameter vectors can produce similar curves. We approximate the forward map with a neural network and use the approximation only inside a bounded search. Figure 1 shows the four steps.

Extraction pipeline diagram

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.

Step 1: build a local simulator bank. For each geometry, we draw 10,000 parameter vectors from a Latin hypercube within estimated bounds, and NGSpice simulates the output and transfer sweeps for each vector. The bank has two uses: it is the supervised training set for the surrogate, and it is a broad initial search in which every candidate’s currents come from the circuit simulator. Because the samples lie near the released cards, the surrogate is trained only on the local extraction region. Each parameter set has 180,000 simulated examples across the 18 geometries.

Step 2: train the forward surrogate. For each parameter set, we train one geometry-conditioned multilayer perceptron. It takes the parameter vector, transistor polarity, length, and width as input and predicts all output and transfer currents together. Predicting all curves at once lets the extraction compare a candidate card across all operating regions in one evaluation. The loss is the sum of a signed-log current term, which emphasizes low-current behavior, and a normalized linear current term, which emphasizes strong-inversion curve shape. The seven- and 43-parameter models use the same architecture and training procedure (Appendix § 10).

Step 3: refine the best bank candidates. For each measured device, several low-error bank samples initialize bounded trust-region least squares [17]. The starts are separated in parameter space so that the result does not depend on a single initialization or local basin. At each iteration, the optimizer compares surrogate-predicted currents with the measurement, and two-point finite differences of the frozen surrogate give the Jacobian. The search can therefore move continuously between sampled cards without NGSpice derivatives and without an NGSpice call for each perturbation.

Step 4: confirm in NGSpice. The surrogate is an approximation and can rank nearby candidates incorrectly. For each device, we re-simulate in NGSpice the leading refined endpoints, their simulator-bank starts, and the anchor (the released card expressed in the search parameterization). We select the candidate with the lowest RRMS under the same fixed scorer as the baseline. The surrogate only proposes refined endpoints, and BSIM4 in NGSpice computes every reported current.

Results

Both searches reduce the in-sample mean device RRMS relative to the re-simulated released cards (Figure 2), from 0.275 to 0.228 with seven parameters and to 0.204 with 43 parameters. The 43-parameter search has the lowest all-device, nMOS, and pMOS means, and a lower RRMS than the seven-parameter search on 16 of 18 devices. The ranking changes on the high-bias subset, where the seven-parameter cards have the lower mean (0.074 vs. 0.076). Appendix § 12 lists the per-device values.

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|V_{GS}| \ge1.48\ \mathrm{V} and the transfer curves at fixed ∣VDS∣≥1.48 V|V_{DS}| \ge1.48\ \mathrm{V}.

The best cards already present in the simulator banks are close to the final scores. The best stored-bank means are 0.2297 (seven parameters) and 0.2068 (43 parameters), compared with final means of 0.2284 and 0.2038. In each parameter set, local finite-difference refinement supplies the selected card for 7 of the 18 devices, and the remaining 11 selected cards are stored NGSpice starts. Under this sampling range and optimization budget, the synthetic bank provides most of the improvement, and the surrogate adds a smaller refinement. Appendix § 10 gives the bank and endpoint scores.

Figure 3 shows where the error occurs within one sweep. It plots the output characteristic of the nMOS L=1 μmL = 1\,\mu\mathrm{m}, W=3 μmW = 3\,\mu\mathrm{m} device at VGS=1.85 VV_{GS} = 1.85\,\mathrm{V}. The 43-parameter card matches the measured saturation current, while the released and seven-parameter cards predict less current. All three cards predict more current than measured below VDS≈0.5 VV_{DS} \approx0.5\,\mathrm{V}, so the fitted cards improve the saturation region but leave the low-VDSV_{DS} mismatch in place.

Figure 3. Output characteristic of the nMOS L=1 μmL = 1\,\mu\mathrm{m}, W=3 μmW = 3\,\mu\mathrm{m} device at VGS=1.85 VV_{GS} = 1.85\,\mathrm{V}. Left: full VDSV_{DS} sweep. Right: VDS≤0.37 VV_{DS} \le0.37\,\mathrm{V}. Lines are NGSpice re-simulations; markers are measurements.

Figure 4 checks whether the aggregate improvement holds curve by curve. Most curves have lower RRMS with the fitted cards, but neither search improves every curve: 33 of 175 curves with seven parameters and 38 of 175 with 43 parameters have materially higher RRMS than with the released cards. The complete overlays in Appendix § 14 show these local changes for all 198 curves.

Figure 4. Per-curve RRMS, released card vs. fitted card. Each panel shows the 156 of 175 scored curves with released and fitted RRMS at most 0.5; the counts in the panel titles include all 175. Red markers and shading mark curves whose fitted RRMS exceeds the released-card RRMS by more than 0.01 and by more than 10%. The flag uses a fixed threshold; no significance test is applied. One flagged seven-parameter curve lies outside the plotted range; all 38 flagged 43-parameter curves are shown.

Discussion

The bank covers the bounded parameter space coarsely, and the surrogate gives local search directions between those samples. NGSpice confirmation keeps surrogate-predicted scores out of the reported results, but surrogate error can still change the optimization path and the shortlist of candidates. The 43-parameter surrogate has a larger approximation error than the seven-parameter surrogate (mean absolute surrogate–NGSpice difference in device RRMS of 0.0030 vs. 0.00035 at the selected cards), yet its confirmed cards have the lower extraction error. The surrogate provides local search evaluations, and the compact model in NGSpice is still used for every final evaluation.

Learning the forward map also avoids the need for a unique target parameter vector for each measured device. The synthetic bank directly supplies parameter vectors and their NGSpice responses for supervised training. During extraction, a candidate card is scored by how well its simulated currents match the measurement, and no assumed true parameter vector is used. The fitted parameters remain non-unique. The parameter values are therefore fit results and are not direct measurements of device physics.

The expanded parameter set gives more directions in which to match the measurements, but the improvement is uneven across bias: its overall mean is lower, while its high-bias mean is slightly higher than that of the seven-parameter fit. The objective averages residuals over complete curves, so improvement on many points can outweigh a larger error in a smaller operating region. The low-VDSV_{\mathrm{DS}} mismatch in Figure 3 is one example, and Figure 4 shows that curve-level regressions also occur on other devices. This experiment does not separate whether these residuals come from the parameter bounds, interactions among parameters, the objective weighting, or limits of the BSIM4 model form at 77 K.

Output and transfer characteristics are fitted together, so each candidate is evaluated across the measured DC response. With equal curve weights, the score measures agreement averaged over all curves. A particular circuit operates over a narrower set of bias points, and a circuit-specific extraction could give more weight to those points. The difference between the overall and high-bias rankings shows that this weighting can change which card is preferred.

Surrogate-guided refinement reduces the mean RRMS by 0.0012 (seven parameters) and 0.0030 (43 parameters) beyond the best card in the simulator bank. The bank provides the broad search and the surrogate allows continuous proposals between those samples. The experiment does not determine whether this is because the bank is already dense near low-error regions, because the local objective has shallow or competing directions, or because surrogate error limits the refinement.

Limitations

The measured curves are used as fitting targets, and the same 18 geometries are used for sensitivity screening and surrogate training. No geometry is held out, so the results measure in-sample extraction for these 18 geometries; prediction for other channel dimensions or temperatures is not tested. The results also depend on one training seed, the chosen parameter bounds, the fixed curve-inclusion rule, and the weighting across curves and devices. Measurement uncertainty and repeatability are not propagated into the fit or the curve-regression flags. The study covers DC behavior for one process, temperature, and corner; dynamic response, noise, variability, reliability, and consistency between separately fitted geometries are not evaluated. The output is a set of fitted DC model cards for the 18 measured geometries.

Conclusion

A forward NGSpice surrogate supports bounded refinement of 77 K BSIM4 cards while preserving standard NGSpice model cards. Across 18 measured SKY130 geometries, in-sample mean RRMS falls from 0.275 for re-simulated released cards to 0.228 (seven parameters) and 0.204 (43 parameters). The seven-parameter cards have the lower high-bias mean, and 33 (seven parameters) and 38 (43 parameters) of the 175 curves have materially higher RRMS than with the released cards. In this setting, the surrogate allows continuous local search without NGSpice in the optimization loop, but selection from the simulator bank accounts for most of the improvement. Final NGSpice confirmation and per-curve evaluation identified these regressions.

Acknowledgments and Disclosure of Funding

We thank Faith Beall for helpful discussions that clarified the Mystic extraction flow used in her earlier work [1], and Ashish Rimal for his guidance on the SKY130 PDK. This material is based upon work supported by the U.S. Department of Energy (DOE), Office of Science, Offices of High Energy Physics and Basic Energy Sciences, as part of the Co-design and Heterogeneous Integration in Microelectronics for Extreme Environments (CHIME) Microelectronics Science Research Center (MSRC) (DOE Lab Announcement No. LAB 24-3320). Fermilab is managed by Fermi Forward Discovery Group, LLC, under Contract No. 89243024CSC000002. This work was also supported by the DOE Office of Science, Advanced Scientific Computing Research, as part of the Transformational Model Consortium (DOE Lab Announcement No. LAB 26-3560). O.S. was supported by a Graduate Instrumentation Research Award (GIRA) through the Coordinating Panel for Advanced Detectors (CPAD), funded by the DOE Office of High Energy Physics.

Reproducibility details

Fixed configuration

Table 1 lists the settings shared by the seven- and 43-parameter experiments. Both use seed 20260818. Training and extraction ran with PyTorch 2.12 on the MPS backend of an Apple M2 Max with 64 GB of unified memory. All final current vectors were generated by conda-forge NGSpice 41 on the CPU. Surrogate training took 1,664.2 s for seven parameters and 1,691.8 s for 43 parameters; extraction with the fixed surrogate took 31.5 s and 235.8 s, respectively. Generating the 43-parameter NGSpice Latin hypercube took 2,145.1 s. The seven-parameter run reused an existing simulator bank, for which no separate generation time was recorded. All reported values come from one seeded training and extraction run per parameter set; no across-seed error bars are reported.

ItemValue
Devices8 nMOS, 10 pMOS; 18 measured geometries
Curves per device5 output, 6 transfer; 2,046 points
Synthetic data10,000 LHC rows per geometry and parameter set
Surrogate inputsparameters, polarity, scaled log LL, log WW
Hidden layers4×10244 \times1024; GELU
Split / batch size80%/10%/10% per device; 512 rows
OptimizerAdamW; initial LR 10−310^{-3}; weight decay 10−510^{-5}
Schedule / epochscosine decay; at most 200; patience 25
Hybrid losssigned-log MSE + linear Huber (β=1\beta= 1), weights 1:1
Geometry scalingpolarity and linearly scaled log LL, log WW in [−1,1][-1, 1]
Extraction starts8 diverse NGSpice LHC vectors
Start separationRMS active-coordinate distance ≥0.35\ge0.35
Least squaresTRF, two-point Jacobian, step 0.02
Search limits100 function evaluations; bounds [−1,1][-1, 1]
Prior10−410^{-4} penalty toward released parameters
NGSpice confirmation4 endpoints, 4 starts, anchor

Table 1. Surrogate and extraction configuration.

For both parameter sets, each device sends four surrogate endpoints, their four stored NGSpice starts, and the anchor to NGSpice, giving nine confirmations after de-duplication. Seven of the 18 selected cards are endpoints and 11 are stored starts. The best stored-bank means are 0.22967 (seven parameters) and 0.20678 (43 parameters), compared with final means of 0.22844 and 0.20380. At the selected cards, the mean absolute surrogate–NGSpice difference in device RRMS is 0.00035 (maximum 0.00158) for seven parameters and 0.00299 (maximum 0.01410) for 43 parameters.

The reported baseline always uses the unmodified released cards. In the 43-parameter layout, writing every parameter at its read-back anchor value gives a mean RRMS of 0.27792. The unmodified cards give 0.27505, with a maximum device-level difference of 0.01111. The anchor is included only as a confirmation candidate and is not selected for any reported result.

Let s(I)=sign⁡(I)log⁡(1+∣I∣/10−9 A)s(I)=\operatorname{sign}(I)\log(1+|I|/10^{-9}\,\mathrm{A}) and zj=[s(Ij)−μj]/σjz_j=[s(I_j)-\mu_j]/\sigma_j, where μj\mu_j and σj\sigma_j are computed pointwise over the synthetic bank. For device dd and current point jj, the linear-current scale ad,ja_{d,j} is the 99th percentile of ∣Ij∣|I_j| within that curve, bounded below by both 10−15 A10^{-15}\,\mathrm{A} and 10−610^{-6} times the largest curve scale of that device. With uniform point weights, the forward loss is

Lfwd=mean⁡j(z^j−zj)2+mean⁡jH1(I^j−Ijad,j)(2)\mathcal{L}_{\mathrm{fwd}}=\operatorname{mean}_{j}(\hat{z}_j-z_j)^2+\operatorname{mean}_{j}H_1\left(\frac{\hat{I}_j-I_j}{a_{d,j}}\right) \tag*{(2)}

where H1H_1 is the smooth-L1L_1 loss with transition 1. The checkpoint with the best validation loss is kept, with patience 25 and a 200-epoch limit; both runs reach the limit. On the 18,000-example synthetic test split, the total loss is 9.14×10−79.14 \times10^{-7} for seven parameters and 1.06×10−31.06 \times10^{-3} for 43 parameters. The signed-log MSE components are 6.37×10−76.37 \times10^{-7} and 9.96×10−49.96 \times10^{-4}, and the linear Huber components are 2.77×10−72.77 \times10^{-7} and 6.70×10−56.70 \times10^{-5}. Because the normalization statistics use the full synthetic bank, these values measure the interpolation error of the surrogate on simulated data.

For active extraction coordinates xA∈[−1,1]DAx_A \in[-1,1]^{D_A}, where AA is the device-specific active set and DAD_A its size, the least-squares residual gives every admitted curve equal weight. The data term is

Linv(xA)=1K∑k=1K1Nk∑i=1Nk(I^k,i(xA)−Ik,iNk−1∑j∣Ik,j∣)2+10−4DA∥xA∥22.(3)\mathcal{L}_{\mathrm{inv}}(x_A)=\frac{1}{K}\sum_{k=1}^{K}\frac{1}{N_k}\sum_{i=1}^{N_k}\left(\frac{\hat{I}_{k,i}(x_A)-I_{k,i}}{N_k^{-1}\sum_j\lvert I_{k,j}\rvert}\right)^2+\frac{10^{-4}}{D_A}\lVert x_A\rVert_2^2. \tag*{(3)}

Curve handling and scoring

The curve order is output sweeps at fixed ∣VGS∣=0.37\lvert V_{GS}\rvert= 0.37, 0.74, 1.11, 1.48, 1.85 V, followed by transfer sweeps at fixed ∣VDS∣=0.01\lvert V_{DS}\rvert= 0.01, 0.37, 0.74, 1.11, 1.48, 1.85 V. The implementation keeps the released pMOS negative sweep direction, device multiplicity, parasitic instance values, and model-bin selection. Measured currents before the last exact zero are set to zero. Leading pMOS points are trimmed at the released device-specific current floors. The exact (IVG,IVD)(I_{VG}, I_{VD}) floors in amperes, in the order of Table 2, are

TypeLL (µm)WW (µm)Released7 param.43 param.
nMOS0.151.60.05950.03510.0380
nMOS0.1970.09670.04360.0346
nMOS0.251.60.09830.04870.0438
nMOS11.60.12830.06370.0648
nMOS130.15990.11020.0657
nMOS81.60.14740.08090.0766
nMOS200.640.12830.11610.1083
nMOS1001000.14030.13240.1317
pMOS0.350.550.72700.63350.6203
pMOS0.351.60.37760.33480.2691
pMOS0.3550.32910.28820.2813
pMOS0.50.420.45600.36700.3161
pMOS0.50.640.34980.32820.3108
pMOS250.19070.17830.1670
pMOS470.27020.22710.1855
pMOS80.840.43400.38130.3343
pMOS81.60.54330.47680.3940
pMOS850.31460.26620.2263
All-device mean0.27510.22840.2038

Table 2.

(10−10,0)(10^{-10}, 0), (6×10−10,0)(6 \times10^{-10}, 0), (6×10−6,10−7)(6 \times10^{-6}, 10^{-7}), (10−7,5×10−8)(10^{-7}, 5 \times10^{-8}), (0,0)(0, 0), (6×10−10,0)(6 \times10^{-10}, 0), (1.2×10−9,0.75)(1.2 \times10^{-9}, 0.75), (1.8×10−9,0)(1.8 \times10^{-9}, 0), (10−9,0)(10^{-9}, 0), and (4.5×10−9,0)(4.5 \times10^{-9}, 0).

For nMOS, a curve is admitted when the trimmed mean ∣Imeas∣\lvert I_{\mathrm{meas}}\rvert exceeds 3×10−73 \times10^{-7} A for output curves or 2×10−72 \times10^{-7} A for transfer curves; the L=W=100 μmL=W=100\ \mu\mathrm{m} device uses 3×10−53 \times10^{-5} A and 9×10−109 \times10^{-10} A. For pMOS, the final simulated magnitude of the released card must exceed 10−1010^{-10} A for output curves or 10−1110^{-11} A for transfer curves. We freeze this released-card inclusion set before fitting. The 23 excluded curves are the ∣VGS∣=0.37\lvert V_{GS}\rvert= 0.37 V output curves of all eight nMOS devices; the ∣VGS∣=0.37\lvert V_{GS}\rvert= 0.37 V output curves of every pMOS device except L=0.35 μmL=0.35\ \mu\mathrm{m}, W=0.55 μmW=0.55\ \mu\mathrm{m}; and the ∣VGS∣=0.74\lvert V_{GS}\rvert= 0.74 V output curves of the pMOS devices with (L,W)=(0.5,0.42)(L,W)=(0.5,0.42), (2,5)(2,5), (4,7)(4,7), (8,0.84)(8,0.84), (8,1.6)(8,1.6), and (8,5) μm(8,5)\ \mu\mathrm{m}. No transfer curve is excluded. Equation (1) is evaluated on the same cleaned and trimmed spans for every method.

Parameter sweep details

To sweep the seven selected parameters, we first added a multiplier to each of them in the room-temperature PDK. Each multiplier is defined as 1 in the global section of the PDK and can be set externally for each sweep and each device. We then chose a range for each nominal multiplier from the expected behavior of the corresponding parameter. The Latin-hypercube script checks each range by confirming that NGSpice produces a valid I–V curve at both ends of the range. The validated ranges are passed to the Latin-hypercube Monte Carlo framework, which generates the sampled multiplier values and sends them to NGSpice to produce the training data. The two previously extracted PDKs are used only as the baseline; their extracted multiplier values were not used to set the ranges, so this procedure does not depend on a prior extraction and can be applied to a PDK that has not been optimized before.

Parameter sets

The original seven-parameter set is

VTH0, U0, NFACTOR, VSAT, DELTA, RDSW, ETA0.

For the expanded set, 115 candidate continuous DC-current parameters were read from the nMOS and pMOS BSIM4 cards. Each candidate was perturbed by a nominal 10% span and simulated on all 11 curves. For a parameter whose released value is zero, the span uses the median nonzero magnitude within that polarity, with the median across both polarities as the fallback. For curve kk, the normalized response to perturbation pp is

Sp,k=RMS⁡(Ip,k−Irel,k)max⁡(mean⁡∣Imeas,k∣,max⁡(10−15 A,10−6Ipeak)).(4)S_{p,k}=\frac{\operatorname{RMS}(I_{p,k}-I_{\mathrm{rel},k})}{\max(\operatorname{mean}\lvert I_{\mathrm{meas},k}\rvert,\max(10^{-15}\ \mathrm{A},10^{-6}I_{\mathrm{peak}))}.} \tag*{(4)}

where IpeakI_{\mathrm{peak}} is the largest measured current magnitude for that device. A parameter is kept if Sp,k≥10−4S_{p,k}\geq10^{-4} for either perturbation direction on any curve and geometry. The original seven parameters are always kept as a nested subset. Non-finite perturbations are retried at 5%, 2.5%, and 1% of the nominal span; the released value and the largest finite perturbation in each available direction define a device-specific, possibly asymmetric interval. Parameters with categorical behavior, geometry-selection roles, capacitance- or noise-only effects, unsafe discontinuities, exact aliases, or no response above the threshold are not kept. Exact per-device bounds are listed in parameter_sensitivity_summary.json. The union over all 18 devices contains the following 43 parameters:

TOXE, TOXM, XJ, NDEP, NSD, DWG, RSH, VTH0, K1, K2, K3, DVT0, DVT1, DVT1W, W0, K3B, VFB, LPE0, LPEB, VSAT, UA, UB, RDSW, PRWG, WR, U0, A0, A2, AGS, B0, B1, VOFF, NFACTOR, CIT, CDSCD, ETA0, DSUB, PCLM, PDIBLC2, DROUT, DELTA, ALPHA0, and BETA0.

A parameter in the all-device union is an input for both polarities. Parameters that are absent from a card, fixed at zero without a safe nonzero scale, or without a non-degenerate safe interval are held at their released value for that device. This leaves 38 active dimensions for nMOS and 39 or 40 for pMOS within a common 43-feature input layout.

Per-device RRMS

Table 2: Per-device RRMS from NGSpice re-simulations. “Released” is the released 77 K parameter vector re-simulated with the same simulator and scorer.

Alternate modeling method

The alternate method has two stages: a forward surrogate and a self-supervised inverse-extraction network. In the first stage, the forward surrogate is trained by supervised regression to map seven SPICE model parameters to a 370-dimensional I–V response. The surrogate has three or four fully connected hidden layers with LeakyReLU activations, L2L_2 weight regularization, and optional dropout. Layer width, dropout rate, and learning rate are selected with Keras Tuner, and training uses the Adam optimizer with an exponentially decaying learning rate.

In the second stage, an inverse network maps a measured I–V response to the seven model parameters. Parameter labels are not used to train this stage. Instead, the predicted parameters pass through the frozen forward surrogate, and the reconstructed I–V curve is compared with the input curve using the mean squared error in log-current space. Gradients through the frozen surrogate update the inverse network while the forward model stays fixed. The forward and inverse networks use the same general dense-network design but are tuned separately.

This result covers only the nMOS device with L=0.15 μmL = 0.15\,\mu\mathrm{m} and W=1.6 μmW = 1.6\,\mu\mathrm{m} at fixed ∣VGS∣=1.48 V\lvert V_{GS}\rvert= 1.48\ \mathrm{V}. Its training set contains approximately 56,000 I–V curves generated with the same Latin-hypercube sampling procedure. It is a single-device comparison and is not included in the 18-device aggregate results.

Bar chart of mean RRMS for the nMOS $L = 0.15\,\mu\mathrm{m}$, $W = 1.6\,\mu\mathrm{m}$ device at fixed $\lvert V_{GS}\rvert = 1.48\ \mathrm{V}$, with bars for Released 77 K card, 7 parameters, 43 parameters, and Alternate 7 param., valued 0.014, 0.022, 0.023, and 0.051

Figure 5. RRMS for the nMOS L=0.15 μmL = 0.15\,\mu\mathrm{m}, W=1.6 μmW = 1.6\,\mu\mathrm{m} device at fixed ∣VGS∣=1.48 V\lvert V_{GS}\rvert= 1.48\ \mathrm{V} for the main and alternate methods.

Complete I–V overlays

The following figures show all 198 measured curves: 11 sweeps for each of the 18 geometries. The frozen scorer admits 175 of these curves, as described in Appendix § 10; the excluded curves are also plotted. Markers are measured data. Dashed, dash-dotted, and solid lines are NGSpice re-simulations of the released, seven-parameter, and 43-parameter cards. Color denotes fixed ∣VGS∣\lvert V_{GS}\rvert in the output panel and fixed ∣VDS∣\lvert V_{DS}\rvert in the transfer panels. Current magnitudes are plotted for both polarities. The linear panels show the strong-inversion shape, and the logarithmic transfer panel shows the subthreshold and near-off regions.

Figure 6. Complete I–V overlays for nMOS devices (1 of 4).

Figure 7. Complete I–V overlays for nMOS devices (2 of 4).

Figure 8. Complete I–V overlays for nMOS devices (3 of 4).

Figure 9. Complete I–V overlays for nMOS devices (4 of 4).

Figure 10. Complete I–V overlays for pMOS devices (1 of 5).

Figure 11. Complete II–VV overlays for pMOS devices (2 of 5).

Figure 12. Complete I–V overlays for pMOS devices (3 of 5).

Figure 13. Complete I–V overlays for pMOS devices (4 of 5).

Figure 14. Complete I–V overlays for pMOS devices (5 of 5).

References

  1. [1]F. Beall, A. Rimal, O. Seidel, Y. Mei, A. D. McDonald, I. Parmaksiz, V. A. Chirayath, J. Asaadi, D. Braga, and J. B. R. Battat. DC Cryogenic Modeling of Open-Source Sky/Water 130 nm MOSFETs at 77 K. Using BSIM4, 2026. URL https://arxiv.org/abs/2604.21625. arXiv:2604.21625v2.
  2. [2]Arnout Beckers, Farzan Jazaeri, and Christian Enz. Cryogenic MOS Transistor Model. IEEE Transactions on Electron Devices, 65(9):3617–3625, 2018. doi: 10.1109/TED.2018.2854701. URL https://doi.org/10.1109/TED.2018.2854701.
  3. [3]Arnout Beckers, Farzan Jazaeri, Heorhii Bohuslavskyy, Louis Hutin, Silvano De Franceschi, and Christian Enz. Characterization and Modeling of 28-nm FDSOI CMOS Technology Down to Cryogenic Temperatures. Solid-State Electronics, 159:106–115, 2019. doi: 10.1016/j.sse.2019.03.033. URL https://doi.org/10.1016/j.sse.2019.03.033.
  4. [4]Arnout Beckers, Farzan Jazaeri, and Christian Enz. Theoretical Limit of Low Temperature Subthreshold Swing in Field-Effect Transistors. IEEE Electron Device Letters, 41(2):276–279, 2020. doi: 10.1109/LED.2019.2963379. URL https://doi.org/10.1109/LED.2019.2963379.
  5. [5]H. Bohuslavskyy, A. G. M. Jansen, S. Barraud, V. Barral, M. Casse, L. Le Guevel, X. Jehl, L. Hutin, B. Bertrand, G. Billiot, G. Pillonnet, F. Arnaud, P. Galy, S. De Franceschi, M. Vinet, and M. Sanquer. Cryogenic Subthreshold Swing Saturation in FD-SOI MOSFETs Described With Band Broadening. IEEE Electron Device Letters, 40(5):784–787, 2019. doi: 10.1109/LED.2019.2903111. URL https://doi.org/10.1109/LED.2019.2903111.
  6. [6]Davide Braga, Shaorui Li, and Farah Fahim. Cryogenic Electronics Development for High-Energy Physics: An Overview of Design Considerations, Benefits, and Unique Challenges. IEEE Solid-State Circuits Magazine, 13(2):36–45, 2021. doi: 10.1109/MSSC.2021.3072804. URL https://doi.org/10.1109/MSSC.2021.3072804.
  7. [7]BSIM Group. BSIM4 4.8.3 MOSFET Model: User’s Manual. University of California, Berkeley, 2025. URL https://bsim.berkeley.edu/models/bsim4/.
  8. [8]Fredo Chavez, Chien-Ting Tung, Ming-Yen Kao, Chenming Hu, Jen-Hao Chen, and Sourabh Khandelwal. Deep Learning-Based I–V Global Parameter Extraction for BSIM-CMG. Solid-State Electronics, 209:108766, 2023. doi: 10.1016/j.sse.2023.108766. URL https://doi.org/10.1016/j.sse.2023.108766.
  9. [9]Kyle Cranmer, Johann Brehmer, and Gilles Louppe. The Frontier of Simulation-Based Inference. Proceedings of the National Academy of Sciences, 117(48):30055–30062, 2020. doi: 10.1073/pnas.1912789117. URL https://doi.org/10.1073/pnas.1912789117.
  10. [10]Alberto Gatti and Filip Tavernier. Cryogenic Small Dimension Effects and Design-Oriented Scalable Compact Modeling of a 65-nm CMOS Technology. IEEE Journal of the Electron Devices Society, 12:369–378, 2024. doi: 10.1109/JEDS.2024.3394167. URL https://doi.org/10.1109/JEDS.2024.3394167.
  11. [11]Olivia Gzamouranis. CryoPDK_Skywater130nm_ML: 77 K SKY130 Measurement, Simulation, and Model-Card Files, 2026. URL https://github.com/ogzamour/CryoPDK_Skywater130nm_ML/tree/39b1e518e25120104225b8fa19f4cfc61a6766b3.
  12. [12]Jan S. Hesthaven and Stefano Ubbiali. Non-Intrusive Reduced Order Modeling of Nonlinear Problems Using Neural Networks. Journal of Computational Physics, 363:55–78, 2018. doi: 10.1016/j.jcp.2018.02.037. URL https://doi.org/10.1016/j.jcp.2018.02.037.
  13. [13]Rosario M. Incandela, Lin Song, Harald Homulle, Edoardo Charbon, Andrei Vladimirescu, and Fabio Sebastiano. Characterization and Compact Modeling of Nanometer CMOS Transistors at Deep-Cryogenic Temperatures. IEEE Journal of the Electron Devices Society, 6:996–1006, 2018. doi: 10.1109/JEDS.2018.2821763. URL https://doi.org/10.1109/JEDS.2018.2821763.
  14. [14]Ming-Yen Kao, Fredo Chavez, Sourabh Khandelwal, and Chenming Hu. Deep Learning-Based BSIM-CMG Parameter Extraction for 10-nm FinFET. IEEE Transactions on Electron Devices, 69(8):4765–4768, 2022. doi: 10.1109/TED.2022.3181536. URL https://doi.org/10.1109/TED.2022.3181536.
  15. [15]Anton Potočnik. How to Scale the Electronic Control Systems of a Quantum Computer. Nature Electronics, 8(1):3–4, 2025. doi: 10.1038/s41928-024-01331-9. URL https://doi.org/10.1038/s41928-024-01331-9.
  16. [16]The ngspice team. ngspice, Version 41, 2023. URL https://sourceforge.net/projects/ngspice/files/ng-spice-rework/old-releases/41/.
  17. [17]Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, Stéfan J. van der Walt, Matthew Brett, Joshua Wilson, K. Jarrod Millman, Nikolay Mayorov, Andrew R. J. Nelson, Eric Jones, Robert Kern, Eric Larson, C. J. Carey, İlhan Polat, Yu Feng, Eric W. Moore, Jake VanderPlas, Denis Laxalde, Josef Perktold, Robert Cimmrman, Ian Henriksen, E. A. Quintero, Charles R. Harris, Anne M. Archibald, António H. Ribeiro, Fabian Pedregosa, Paul van Mulbregt, and SciPy 1.0 Contributors. SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python. Nature Methods, 17(3):261–272, 2020. doi: 10.1038/s41592-019-0686-2. URL https://doi.org/10.1038/s41592-019-0686-2.

Paper details

vlnc:W_mqrh26y4
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 3

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents