Oursdesvosges/coilstellaration-phys-resnet-loopfree-universal

PhysResNet-LoopFree v2.3 (Universal)

Single-file self-contained CPU ONNX submission (model.onnx, 36.54 MiB <= 50 MiB, opset 17, CPUExecutionProvider, no external .data sidecars) for Proxima Fusion's CoilStellaration benchmark (bench_version=1.0, dataset revision f505cc890ff2f5cf24a6d68f4b5156d0a3fecb5c, arXiv:2611.12689). Supports all 17 (NFP, N_c) topologies (NFP in {2..5}, N_c in {4..8}) across fixed_shape and variable_shape.

Architecture Summary

• Model Family: Universal loop-free physics ResNet + lightweight Dual-Init selector + If-gated unrolled analytic-Jacobian Gauss-Newton step (Loop = 0 in the exported ONNX graph).
• Design: Stage-E multi-topology PhysResNet (2 differentiable CorrectorStages), followed by a 16x16x64 (pdist_cov=1.2) Dual-Init selector (x_net vs. x_tmpl) and an If-gated unrolled analytic-Jacobian Gauss-Newton stage (14x14x48 LM grid, cg_iters = 14, col_eps = 1e-2) that executes only when the warm-start fails the discrete gate.
• Training & Tuning Discipline: Selected strictly on val_holdout_suite.pkl (N = 360 fixed + N = 450 variable boundaries sampled from split="train"). Zero split="eval" data touched training or hyperparameter selection.

Official split="eval" Results (bench_version=1.0, overall_score=0.5160)

Track / Stratum	N	Mean Score	Median	p10	Feas >= 0.9	Strict Feas
fixed_shape	2,771	0.5176	0.4149	0.0237	36.45%	33.31%
fixed_shape/loose	1,026	0.9458	1.0000	0.8086	86.65%	81.29%
fixed_shape/medium	861	0.4049	0.3187	0.0536	13.94%	10.22%
fixed_shape/tight	884	0.1303	0.0800	0.0049	0.11%	0.11%
variable_shape	8,058	0.5144	0.4328	0.0257	33.36%	29.32%
variable_shape/loose	2,697	0.9294	1.0000	0.7149	82.28%	74.94%
variable_shape/medium	2,660	0.4703	0.4126	0.0692	17.26%	12.78%
variable_shape/tight	2,701	0.1434	0.0871	0.0041	0.37%	0.07%

8-vCPU ONNX Runtime Latency (CPUExecutionProvider, intra_op=8, N = 200 / track)

Measured on an idle 8-vCPU allocation (separate from the 384-worker Slurm scoring array):

• fixed_shape: mean = 125.9 ms, median = 123.5 ms, p90 = 166.7 ms, max = 249.3 ms
• variable_shape: mean = 141.0 ms, median = 128.5 ms, p90 = 206.6 ms, max = 353.1 ms

Validation Holdout (split="train", N = 360 / 450)

Track	N	Mean Score	p10	Feas >= 0.9	Strict Feas
fixed_shape	360	0.4808	0.0276	31.67%	28.33%
variable_shape	450	0.5263	0.0319	36.00%	31.33%

Quick Usage (onnxruntime)

import numpy as np
import onnxruntime as ort
from huggingface_hub import hf_hub_download

path = hf_hub_download("Oursdesvosges/coilstellaration-phys-resnet-loopfree-universal", "model.onnx")
sess = ort.InferenceSession(path, providers=["CPUExecutionProvider"])

r_cos = np.zeros((5, 9), dtype=np.float32)
z_sin = np.zeros((5, 9), dtype=np.float32)
r_cos[0, 4], r_cos[1, 4], z_sin[1, 4] = 1.0, 0.15, 0.15

coil_x_n, coil_y_n, coil_z_n, currents = sess.run(
    ["coil_x_n", "coil_y_n", "coil_z_n", "currents"],
    {
        "boundary_r_cos": r_cos,
        "boundary_z_sin": z_sin,
        "requirement_metrics": np.array([0.4, 0.8, 3.0, 0.05], dtype=np.float32),
        "n_field_periods": np.asarray(3, dtype=np.int64),
        "n_coils_per_half_period": np.asarray(5, dtype=np.int64),
    },
)