Foundational ML Model for Physics
Levenson-Falk Lab
QCap
QCap predicts the full Maxwell capacitance matrix of any planar superconducting circuit layout straight from its GDS file, in about a second on a laptop instead of minutes to hours of Ansys Q3D simulation. It reads every metal piece of the layout as a node of a graph, so it works for any layout made with any tool, and it never merges nets: an N-conductor layout gets its full N x N matrix.
The model is physics first. A fast thin-film method-of-moments solve gives a first estimate of how every pair of conductors couples, and a graph neural network, trained on about 40,000 Ansys Q3D simulations, learns the correction. On a family of layouts that no part of the model saw in training, QCap’s median error is 1.05 % (MAE 0.132 fF, RMSE 0.188 fF).
QCap is open source in SQuADDS (pip install "squadds[ml]"), with the dataset, the models, a browser demo and a hosted API. The preprint is coming very soon.
Resources
Paper Docsite: https://lfl-lab.github.io/SQuADDS/qcap/, the paper section by section with short videos
Browser Demo: https://huggingface.co/spaces/SQuADDS/planar-capacitance
Models: https://huggingface.co/SQuADDS/QCap
Dataset: https://huggingface.co/datasets/SQuADDS/SQuADDS_PlanarCapacitance
Docs and Tutorials: ML Models guide and Tutorials 11-16