Card: Fast is an improved release of the paper's fast model; each tier quotes its own in-sample fit and B5, the paper's held-out numbers are labelled as the paper's version; physical-check warning
Browse files
README.md
CHANGED
|
@@ -6,7 +6,7 @@ tags: [superconducting-qubits, capacitance, electrostatics, gds, graph-neural-ne
|
|
| 6 |
|
| 7 |
# Planar capacitance: flagship models (SQuADDS)
|
| 8 |
|
| 9 |
-
GDS layout -> full Maxwell capacitance matrix of superconducting-circuit components. Networks trained on all 45,771 on-stack Q3D layouts
|
| 10 |
|
| 11 |
```python
|
| 12 |
from squadds.ml.capacitance import CapacitancePredictor, load_model
|
|
@@ -17,39 +17,41 @@ res.names, res.matrix_fF, res.to_sim_results()
|
|
| 17 |
| tier | folder | what |
|
| 18 |
|---|---|---|
|
| 19 |
| **Accurate** (default) | `accurate/` | Two thin-film method-of-moments solves (corrected prior with substrate far field, 1.5 and 3 panels per gap), Richardson-extrapolated, corrected by a 3-seed ensemble of the paper model's networks (89,858 parameters per seed, bounded residual) with a learned row-sum closure: the complete Maxwell matrix, self terms included. Networks retrained on all 45,771 on-stack Q3D layouts (2026-10-20, on the corrected prior with substrate far field). |
|
| 20 |
-
| Fast | `fast/` |
|
| 21 |
|
| 22 |
-
> **Warning (Fast).** Use Fast only for quick screening of the component families it was trained on (interdigital, paddle, concentric, star, multipad star, Xmon, TransmonPocket, shielded concentric, two-net, random, procedurally generated and synthetic campaign layouts). On new designs of a trained family (B5) it was worse than the physics alone, 3.175 % (MAE 0.583 / RMSE 0.824 fF) against 1.765 % (0.275 / 0.376 fF), and
|
| 23 |
|
| 24 |
## Accurate
|
| 25 |
|
| 26 |
-
|
| 27 |
|
| 28 |
-
|
|
| 29 |
|---|---|
|
| 30 |
-
|
|
| 31 |
-
|
|
| 32 |
-
| paper recipe, test zero-shot macro (1,878) | 0.587 % (MAE 1.092 fF / RMSE 1.152 fF) |
|
| 33 |
-
| paper recipe, held-out family (4,071) | 0.408 % (MAE 0.049 fF / RMSE 0.059 fF) |
|
| 34 |
-
| B5 fresh check (39 layouts) | 1.05 % (MAE 0.132 fF / RMSE 0.188 fF) |
|
| 35 |
| B5: physics alone C^R = R(1.5, 3) | 1.765 % (MAE 0.275 fF / RMSE 0.376 fF) |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
|
| 37 |
Time: x86 cluster node, 2 CPUs: 0.845 / 5.96 s in family, 1.75 / 18.1 s zero-shot (median / p90; x86 cluster, 2 CPUs); B5 0.99 s median (1.14 s on the 2026-10-20 re-run, Xeon E5-2640 v4). HF Space (2 vCPU): interdigital 0.47 s, xmon 0.91 s, 1-qubit lattice (N 15) 19.0 s, 16-conductor layout 9.9 s, 9-qubit lattice (N 67) 154 s (server time, one request, Docker build 2026-10-20; a repeated request is answered from the cache in ~0.01 s). SQuADDS package (median per layout, cold caches, warm numba; 2 / 8 threads): GH200 (aarch64) N<=4 1.88 / 0.98 s, N 5-9 6.41 / 3.42 s, N 10-20 23.0 / 10.5 s, N 35 22.0 / 9.5 s; x86 Xeon Silver 4116 N<=4 3.76 / 3.10 s, N 5-9 13.35 / 9.60 s, N 10-20 46.2 / 29.2 s, N 35 44.6 / 24.9 s. The first call of a session also loads the model and compiles the solver (about 40 s once); a repeated layout is answered from the cache in milliseconds.
|
| 38 |
|
| 39 |
## Fast
|
| 40 |
|
| 41 |
-
|
| 42 |
|
| 43 |
-
|
|
| 44 |
|---|---|
|
| 45 |
-
|
|
| 46 |
-
|
|
| 47 |
-
|
|
| 48 |
-
|
|
| 49 |
-
|
|
| 50 |
-
|
|
| 51 |
-
|
|
| 52 |
-
|
|
|
|
|
| 53 |
|
| 54 |
Time: x86 cluster node, 2 CPUs: 0.091 / 0.274 s in family, 0.30 / 2.45 s zero-shot (median / p90; x86 cluster, 2 CPUs); B5 0.24 s median. HF Space (2 vCPU): interdigital 0.036 s, xmon 0.11 s, 1-qubit lattice (N 15) 0.82 s, 16-conductor layout 0.49 s, 9-qubit lattice (N 67) 3.9 s (server time, one request, Docker build 2026-10-20; a repeated request is answered from the cache in ~0.01 s). SQuADDS package (median per layout, cold caches, warm numba; 2 / 8 threads): GH200 (aarch64) N<=4 0.093 / 0.084 s, N 5-9 0.25 / 0.20 s, N 10-20 0.32 / 0.34 s, N 35 0.82 / 0.59 s; x86 Xeon Silver 4116 N<=4 0.17 / 0.22 s, N 5-9 0.33 / 0.30 s, N 10-20 0.53 / 0.56 s, N 35 2.21 / 1.96 s. The first call of a session also loads the model (and the embedding code compiles once).
|
| 55 |
|
|
|
|
| 6 |
|
| 7 |
# Planar capacitance: flagship models (SQuADDS)
|
| 8 |
|
| 9 |
+
GDS layout -> full Maxwell capacitance matrix of superconducting-circuit components. Networks trained on all 45,771 on-stack Q3D layouts, the paper's test split included, so these models have no held-out score of their own. Each tier lists its own in-sample fit (a fit to data it was trained on, not a generalisation score), one fresh check (B5: 39 new layouts of one generated family, 3-4 conductors, never seen by any component) and, labelled "paper's version", the held-out numbers of the paper's model of that tier, trained on the 35,354 TRAIN layouts only. Those numbers belong to the paper's versions, not to these models. For benchmarks use the paper model (`SQuADDS/PlanarCapacitance-paper`).
|
| 10 |
|
| 11 |
```python
|
| 12 |
from squadds.ml.capacitance import CapacitancePredictor, load_model
|
|
|
|
| 17 |
| tier | folder | what |
|
| 18 |
|---|---|---|
|
| 19 |
| **Accurate** (default) | `accurate/` | Two thin-film method-of-moments solves (corrected prior with substrate far field, 1.5 and 3 panels per gap), Richardson-extrapolated, corrected by a 3-seed ensemble of the paper model's networks (89,858 parameters per seed, bounded residual) with a learned row-sum closure: the complete Maxwell matrix, self terms included. Networks retrained on all 45,771 on-stack Q3D layouts (2026-10-20, on the corrected prior with substrate far field). |
|
| 20 |
+
| Fast | `fast/` | An improved release of the paper's fast model, same architecture: a learned surrogate (1.02M parameters) predicts the two uncorrected-prior solves at 1.5 and 3 panels per gap and residual networks correct them. The surrogate is trained on more generated layouts (the prior of every on-stack layout plus 432,697 generated layouts labelled by the prior alone, against 385,593 for the paper's) and the networks are retrained on every labelled design (all 45,771 on-stack layouts, the paper's test split included). No field solve: the time is the geometric embedding. |
|
| 21 |
|
| 22 |
+
> **Warning (Fast).** Use Fast only for quick screening of the component families it was trained on (interdigital, paddle, concentric, star, multipad star, Xmon, TransmonPocket, shielded concentric, two-net, random, procedurally generated and synthetic campaign layouts). On new designs of a trained family (B5) it was worse than the physics alone, 3.175 % (MAE 0.583 / RMSE 0.824 fF) against 1.765 % (0.275 / 0.376 fF), and the paper's version of it was worse than the physics alone on a held-out family too, 3.11 % (0.533 / 0.633 fF) against 1.55 % (0.166 / 0.205 fF). Use Accurate for anything that matters. Every result carries its physical checks (symmetry, off-diagonal signs, positive-definite signal block); when one fails, the result says so in its warnings (and SQuADDS raises a `UserWarning`): do not use that matrix.
|
| 23 |
|
| 24 |
## Accurate
|
| 25 |
|
| 26 |
+
Paper's version: the paper's model (same architecture and prior stack at inference; its networks were trained on the 35,354 TRAIN layouts only, on the uncorrected prior; retrained on the corrected prior, as here, they score validation 0.218 % (0.0774 / 0.0888 fF)). Its held-out numbers below belong to that version, not to this release.
|
| 27 |
|
| 28 |
+
| set | median % (MAE / RMSE) |
|
| 29 |
|---|---|
|
| 30 |
+
| this model, in-sample (all 45,771 training layouts; it was trained on them) | 0.207 % (MAE 0.0679 fF / RMSE 0.077 fF) |
|
| 31 |
+
| this model, B5 fresh check (39 layouts, never seen) | 1.05 % (MAE 0.132 fF / RMSE 0.188 fF) |
|
|
|
|
|
|
|
|
|
|
| 32 |
| B5: physics alone C^R = R(1.5, 3) | 1.765 % (MAE 0.275 fF / RMSE 0.376 fF) |
|
| 33 |
+
| paper's version, validation (4,608 layouts) | 0.229 % (MAE 0.0674 fF / RMSE 0.0792 fF) |
|
| 34 |
+
| paper's version, test in family (4,009) | 0.177 % (MAE 0.0203 fF / RMSE 0.025 fF) |
|
| 35 |
+
| paper's version, test zero-shot macro (1,878) | 0.587 % (MAE 1.092 fF / RMSE 1.152 fF) |
|
| 36 |
+
| paper's version, held-out family (4,071) | 0.408 % (MAE 0.049 fF / RMSE 0.059 fF) |
|
| 37 |
|
| 38 |
Time: x86 cluster node, 2 CPUs: 0.845 / 5.96 s in family, 1.75 / 18.1 s zero-shot (median / p90; x86 cluster, 2 CPUs); B5 0.99 s median (1.14 s on the 2026-10-20 re-run, Xeon E5-2640 v4). HF Space (2 vCPU): interdigital 0.47 s, xmon 0.91 s, 1-qubit lattice (N 15) 19.0 s, 16-conductor layout 9.9 s, 9-qubit lattice (N 67) 154 s (server time, one request, Docker build 2026-10-20; a repeated request is answered from the cache in ~0.01 s). SQuADDS package (median per layout, cold caches, warm numba; 2 / 8 threads): GH200 (aarch64) N<=4 1.88 / 0.98 s, N 5-9 6.41 / 3.42 s, N 10-20 23.0 / 10.5 s, N 35 22.0 / 9.5 s; x86 Xeon Silver 4116 N<=4 3.76 / 3.10 s, N 5-9 13.35 / 9.60 s, N 10-20 46.2 / 29.2 s, N 35 44.6 / 24.9 s. The first call of a session also loads the model and compiles the solver (about 40 s once); a repeated layout is answered from the cache in milliseconds.
|
| 39 |
|
| 40 |
## Fast
|
| 41 |
|
| 42 |
+
This is an improved release of the paper's fast model: the same architecture, a surrogate trained on more generated layouts and networks retrained on every labelled design, the paper's test split included. Paper's version: the surrogate t100_norm (385,593 generated layouts + the TRAIN geometry) with networks trained on the 35,354 TRAIN layouts only. Its held-out numbers below belong to that version, not to this release.
|
| 43 |
|
| 44 |
+
| set | median % (MAE / RMSE) |
|
| 45 |
|---|---|
|
| 46 |
+
| this model, in-sample (all 45,771 training layouts; it was trained on them) | 0.671 % (MAE 0.353 fF / RMSE 0.395 fF) |
|
| 47 |
+
| this model, B5 fresh check (39 layouts, never seen) | 3.175 % (MAE 0.583 fF / RMSE 0.824 fF) |
|
| 48 |
+
| B5: physics alone C^R = R(1.5, 3) (Fast is behind the physics here) | 1.765 % (MAE 0.275 fF / RMSE 0.376 fF) |
|
| 49 |
+
| this surrogate recipe trained without the validation layouts, with the paper's TRAIN-only networks, validation (4,608, held out) | 0.722 % (MAE 0.387 fF / RMSE 0.436 fF) |
|
| 50 |
+
| paper's version, validation (4,608 layouts) | 0.846 % (MAE 0.491 fF / RMSE 0.549 fF) |
|
| 51 |
+
| paper's version, test in family (4,009) | 0.63 % (MAE 0.113 fF / RMSE 0.1414 fF) |
|
| 52 |
+
| paper's version, test zero-shot macro (1,878) | 3.576 % (MAE 4.608 fF / RMSE 5.058 fF) |
|
| 53 |
+
| paper's version, held-out family (4,071) | 3.113 % (MAE 0.533 fF / RMSE 0.633 fF) |
|
| 54 |
+
| paper's version, held-out family: physics alone | 1.554 % (MAE 0.166 fF / RMSE 0.205 fF) |
|
| 55 |
|
| 56 |
Time: x86 cluster node, 2 CPUs: 0.091 / 0.274 s in family, 0.30 / 2.45 s zero-shot (median / p90; x86 cluster, 2 CPUs); B5 0.24 s median. HF Space (2 vCPU): interdigital 0.036 s, xmon 0.11 s, 1-qubit lattice (N 15) 0.82 s, 16-conductor layout 0.49 s, 9-qubit lattice (N 67) 3.9 s (server time, one request, Docker build 2026-10-20; a repeated request is answered from the cache in ~0.01 s). SQuADDS package (median per layout, cold caches, warm numba; 2 / 8 threads): GH200 (aarch64) N<=4 0.093 / 0.084 s, N 5-9 0.25 / 0.20 s, N 10-20 0.32 / 0.34 s, N 35 0.82 / 0.59 s; x86 Xeon Silver 4116 N<=4 0.17 / 0.22 s, N 5-9 0.33 / 0.30 s, N 10-20 0.53 / 0.56 s, N 35 2.21 / 1.96 s. The first call of a session also loads the model (and the embedding code compiles once).
|
| 57 |
|