Part of SQuADDS
QCap: our model trained on all of our Q3D data
The model of our paper QCap: a physics-corrected graph neural network for the capacitance matrix of any planar superconducting device layout, trained on all 45,771 designs of our Q3D dataset, the test designs included. It reads a GDS layout and returns the full Maxwell capacitance matrix of every conductor in it. Because it has seen the test designs, it has no held-out score of its own. Its held-out accuracy, from the same model trained without the test designs, is in the paper.
- Paper page: lfl-lab.github.io/SQuADDS/ml-paper, with the explainer videos and the tutorials
- Paper: lfl-lab.github.io/SQuADDS/ml-paper/paper
- Code: github.com/LFL-Lab/SQuADDS
- Data: Q3D dataset and synthetic dataset, part of the SQuADDS datasets
- Models: QCap models (QCap, QCap-Lite and QCap-Fast), part of the SQuADDS models
- API: SQuADDS ML inference API
- Point of contact: Sadman Ahmed Shanto
Usage
from squadds.ml.capacitance import CapacitancePredictor, load_model
res = CapacitancePredictor(load_model("QCap")).predict("chip.gds") # QCap
fast = CapacitancePredictor(load_model("QCap-Fast")).predict("chip.gds") # QCap-Fast
res.names, res.matrix_fF, res.to_sim_results()
Hosted API
The SQuADDS ML inference API serves these models over HTTP as qcap (QCap) and qcap-fast (QCap-Fast).
GET /v1/models/qcapdescribes the model, with its inputs, limits and an example.GET /v1/models/qcap/schema?format=toolgives a ready tool definition for AI agents.POST /v1/models/qcap/predicttakes the GDS file as the multipart fieldfile, or base64 encoded ininputs.file_base64of a JSON body. The optional inputs areground_layerandconductor_layer(1/0and1/10by default),eps_randthickness_um(11.45 and 350 µm), andground_pieceswhen ground and conductors share a layer.
curl -X POST https://squadds-squadds-ml-inference-api.hf.space/v1/models/qcap/predict \
-F file=@chip.gds -F 'inputs={"ground_layer": "1/0", "conductor_layer": "1/10"}'
from squadds.ml import client
r = client.predict("qcap", file="chip.gds")
r["outputs"]["names"], r["outputs"]["matrix_fF"]
The response for the interdigital example layout:
{
"model_id": "qcap",
"outputs": {
"names": ["G", "C1", "C2"],
"matrix_fF": [[31.727, -4.264, -5.711], [-4.264, 6.598, -2.285], [-5.711, -2.285, 8.045]],
"units": "fF",
"convention": "Maxwell matrix, ground first (row and column 0), nets never merged"
}
}
The API takes up to 70 conductors plus ground and 20 MB per file. It runs one prediction at a time on two shared vCPUs, so large layouts take minutes. The full list of hosted models is at GET /v1/models.
Models
This repository holds all three models in qcap/, qcap-lite/ and qcap-fast/. SQuADDS loads QCap and QCap-Fast.
| model | how it works | error on 39 new layouts, % (MAE / RMSE, fF) | median time |
|---|---|---|---|
| QCap (default) | the extrapolated prior from two coarse method-of-moments solves, corrected by the networks | 1.05 (0.132 / 0.188) | 1.1 s |
| QCap-Lite | the same recipe on two coarser solves | 2.044 (0.258 / 0.348) | 0.56 s |
| QCap-Fast | a learned prior replaces the solves, then the networks correct it | 3.175 (0.583 / 0.824) | 0.24 s |
The 39 layouts are generated rings with a neighbour (3 or 4 conductors) that no part of any model saw in training. On them the extrapolated prior alone scores 1.765 % (MAE 0.275 / RMSE 0.376 fF). Times are on two CPUs of an Intel Xeon E5-2640 v4 cluster node.
QCap-Fast is for quick screening of the component families it was trained on. Use QCap otherwise.
Conventions
- Capacitance is in fF. Ground is row and column 0, and nets are never merged.
- Trained on a 350 µm silicon substrate (eps_r 11.45) with zero-thickness metal. QCap takes other substrates with
predict(..., eps_r=..., thickness_um=...), as in Sec. IV C of the paper. QCap-Fast only rescales by permittivity there and warns. - Ground on GDS layer 1/0 and conductors on 1/10 by default. Every result carries its physical checks, and a failed check is raised as a warning.
Citation
If you use these data or models, please cite our paper and SQuADDS.
@misc{Shanto2026planarcapacitance,
title = {{QCap}: a physics-corrected graph neural network for the capacitance matrix of any planar superconducting device layout},
author = {Shanto, Sadman Ahmed and Das, Saikat and Levenson-Falk, Eli},
year = {2026},
url = {https://lfl-lab.github.io/SQuADDS/ml-paper/paper/}
}
@article{Shanto2024squaddsvalidated,
doi = {10.22331/q-2024-09-09-1465},
url = {https://doi.org/10.22331/q-2024-09-09-1465},
title = {{SQ}u{ADDS}: {A} validated design database and simulation workflow for superconducting qubit design},
author = {Shanto, Sadman and Kuo, Andre and Miyamoto, Clark and Zhang, Haimeng and Maurya, Vivek and Vlachos, Evangelos and Hecht, Malida and Shum, Chung Wa and Levenson-Falk, Eli},
journal = {{Quantum}},
issn = {2521-327X},
publisher = {{Verein zur F{\"{o}}rderung des Open Access Publizierens in den Quantenwissenschaften}},
volume = {8},
pages = {1465},
month = sep,
year = {2024}
}
Licence
MIT.
Files
SQuADDS checks every file against the manifest in the package (squadds/ml/capacitance/manifest.json).
| file | sha256 |
|---|---|
qcap/acc_v2m_sf_s0.pt |
cc69c5818f5028c26177159b88239116520abbb290743d6a6c091657e8dbf1c5 |
qcap/acc_v2m_sf_s1.pt |
2f88606dbec5c599550d088213a4c37af3b3ad5134ebad2e60519fb534553a57 |
qcap/acc_v2m_sf_s2.pt |
d13bafc6b2eb7cadcd97a387d791c0210976db1a9d618f9acc1b614e3701daed |
qcap-lite/bal_v2m_sf_s0.pt |
aa742ab93b1e3ffcd86a753bda5cc4a68812acbe8140a6b0e311c11d1f36b4b2 |
qcap-lite/bal_v2m_sf_s1.pt |
486466473eba61c090f4e3fa7988bf880fc183570c5d02bd78f4c1356691f926 |
qcap-lite/bal_v2m_sf_s2.pt |
1b55ded04fd624585eeeccf2ab6e4a8e9354a9b96694131e1737dde43ad3ff58 |
qcap-fast/acc_v1_s0.pt |
9dff20ade26619e7713c2d145dada65e4554a623780290673cfe9452cf9b64ba |
qcap-fast/acc_v1_s1.pt |
f1ad2ca065dd650ccabcad73a3c6ad2ac1042f4f68bb828893bbdecc47411ffe |
qcap-fast/acc_v1_s2.pt |
8cbada211724df5cc05a7b4fdd232607de047b248a43993d3844bd68ff8060c4 |
qcap-fast/sall_all_s0.pt |
132d6f389eda798d31bd5654a96e24cce18fa750134a9fb2a9725eafed339dac |