PhysResNet-LoopFree v3.1 (Universal)

Single-file self-contained ONNX submission (phys_resnet_loopfree_universal.onnx, 24.66 MiB <= 50 MiB, opset 17, CPU CPUExecutionProvider, zero external .data files) for the Proxima Fusion CoilStellaration benchmark (revision f505cc890ff2f5cf24a6d68f4b5156d0a3fecb5c).

Architecture Summary

  • Model Family: Family A v3.1: 0-Loop Universal Neural-REGCOIL + Unrolled Multi-Mu CG Augmented-Lagrangian
  • Architecture: Universal multi-topology Stage-D PhysResNet (NFP in {2..5}, N_C in {4..8}, 17 topologies, 3 CorrectorStages) fused with ConformalAtlasInit, 3-lambda NeuralRegcoilLayer, top-2 warm-start polish (1 unrolled step), and 9 straight-line unrolled ALBodyStepDyn Multi-Mu fp64 CG stages (K=8, 176 staggered + 144 TopK active surface points from a 28x28 pool, lm_n_coil_pts=128, cg_iters=24, col_eps=1e-2, mu_min=8e-4, hi_reg_scale=5e-3, rf_quad_weight=0.0, 42x42x192 verification grid) exported as a single loop-free ONNX graph (Loop = 0, opset 17).
  • Zero-Leakage Selection: Selected strictly on val_holdout_suite.pkl (split="train" held-out validation, zero split="eval" leakage) as Stage D + Neural-REGCOIL + Top-2 Polish + 9-Stage Unrolled Multi-Mu CG AL (lfu_m2.onnx, 0 Loop ops).

Official split="eval" Results (bench_version=1.0, overall_score=0.8827)

Track / Stratum N Mean Score Median p10 Feas >= 0.9 Strict Feas 8-vCPU / Shard Latency
fixed_shape 2,771 0.8842 1.0000 0.5100 77.19% 70.95% 736.0 ms
fixed_shape/loose 1,026 0.9917 1.0000 1.0000 97.56% 96.39% 2311.2 ms
fixed_shape/medium 861 0.8729 1.0000 0.4471 73.52% 66.20% 5939.4 ms
fixed_shape/tight 884 0.7705 0.9634 0.1658 57.13% 46.04% 6073.6 ms
variable_shape 8,058 0.8813 1.0000 0.4876 76.36% 69.53% 839.0 ms
variable_shape/loose 2,697 0.9905 1.0000 1.0000 97.37% 96.03% 2233.5 ms
variable_shape/medium 2,660 0.8751 1.0000 0.4670 74.85% 67.11% 4632.2 ms
variable_shape/tight 2,701 0.7783 0.9627 0.2149 56.87% 45.46% 5686.3 ms

Zero-Leakage Validation Holdout (val_holdout_suite.pkl)

Track N Mean Score p10 Feas >= 0.9 Strict Feas
fixed_shape 360 0.8872 0.4899 75.56% 70.83%
variable_shape 450 0.8919 0.5763 74.22% 66.00%

ONNX Graph Inputs & Outputs

Compatible directly with the proxima-fusion/coilstellaration leaderboard ONNX runner (inputs: boundary_r_cos, boundary_z_sin, requirement_metrics, n_field_periods, n_coils_per_half_period; outputs: coil_x_n, coil_y_n, coil_z_n, currents).

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Dataset used to train Oursdesvosges/coilstellaration-phys-resnet-loopfree-universal