PhysResNet-LoopFree v3.1 (Fixed-Shape)

Single-file self-contained ONNX submission (phys_resnet_loopfree_fixed.onnx, 25.07 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 Neural-REGCOIL + Unrolled Multi-Mu CG Augmented-Lagrangian
  • Architecture: Stage-E FiLM-conditioned Residual Network (PhysResNetBackbone + 3 differentiable CorrectorStages) fused with ConformalAtlasInit, 3-lambda NeuralRegcoilLayer, top-2 warm-start polish (1 unrolled step), and 10 straight-line unrolled ALBodyStepDyn Multi-Mu fp64 CG stages (K=8, 200 staggered + 160 TopK active surface points from a 30x30 pool, lm_n_coil_pts=128, cg_iters=28, col_eps=1e-2, mu_min=8e-4, hi_reg_scale=5e-3, unsigned RF residual) in a 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 E + Neural-REGCOIL + Dual-Init + 10-Stage Unrolled Multi-Mu CG AL (lff_m1.onnx, 0 Loop ops).

Official split="eval" Results (bench_version=1.0, overall_score=0.9428)

Track / Stratum N Mean Score Median p10 Feas >= 0.9 Strict Feas 8-vCPU / Shard Latency
fixed_shape 2,771 0.9428 1.0000 0.8135 87.01% 82.35% 771.5 ms
fixed_shape/loose 1,026 0.9919 1.0000 1.0000 98.05% 97.27% 2347.5 ms
fixed_shape/medium 861 0.9266 1.0000 0.7332 83.97% 78.75% 6256.6 ms
fixed_shape/tight 884 0.9016 1.0000 0.6129 77.15% 68.55% 6823.8 ms

Zero-Leakage Validation Holdout (val_holdout_suite.pkl)

Track N Mean Score p10 Feas >= 0.9 Strict Feas
fixed_shape 360 0.9524 0.8455 87.78% 83.61%

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-fixed