Skip to content

RESULTS.md

Latest commit

 

History

History
193 lines (158 loc) · 7.66 KB

RESULTS.md

File metadata and controls

193 lines (158 loc) · 7.66 KB

SCPN Fusion Core — Benchmark Results (v3.9.3)

Auto-generated by validation/collect_results.py on 2026-03-09 19:04 UTC. Re-run the script to refresh these numbers on your hardware.

Environment

  • CPU: Intel64 Family 6 Model 167 Stepping 1, GenuineIntel
  • Architecture: AMD64
  • OS: Windows-11-10.0.26200-SP0
  • Python: 3.12.5
  • NumPy: 1.26.4
  • RAM: 31.8 GB
  • Version: 3.9.3
  • Generated: 2026-03-09 19:04 UTC
  • Wall-clock: 55s

Equilibrium & Transport

Metric Value Unit Notes
3D Force-Balance initial residual 3.8002e+05 — Spectral variational method
3D Force-Balance final residual 1.0706e+05 — After 20 iterations
3D Force-Balance reduction factor 3.5× — initial / final
Neural Equilibrium inference (mean) 1.05 ms PCA+MLP surrogate on 129x129 grid
Neural Equilibrium inference (P95) 2.69 ms 129x129 grid

QLKNN Neural Transport Surrogate

Metric Value Unit Notes
Test relative L2 0.0943 — Hard-fail gate < 0.25
Val relative L2 0.0954 —
Train relative L2 0.0917 — val/train = 1.04
Best val MSE 0.002566 —
Architecture 1024×512×256 — MLP hidden dims
Epochs 911 — Early-stopped
Training time 3.6 h GPU
Data source QLKNN-10D (Zenodo DOI 10.5281/zenodo.3497066) —
Backup test relative L2 0.0949 — 512×256×128 architecture

Confinement Scaling (ITPA H-mode)

Metric Value Unit Notes
Machines validated 53 — ITER, JET, DIII-D, ASDEX-U, C-Mod, JT-60U, NSTX, MAST, KSTAR, EAST, SPARC, ARC, TFTR, WEST, TCV, HL-2A, HL-2M, COMPASS, JT-60SA, SST-1, Aditya-U, Globus-M2, NSTX-U, MAST-U
tau_E RMSE 0.0969 s
tau_E relative RMSE 50.1 %
H98 RMSE 0.2954 —
ITER_15MA_baseline τ_E error -1.0 % τ_pred=3.664 s
SPARC_V2C τ_E error 2.9 % τ_pred=0.793 s
β_N RMSE 0.1731 —
ITER_15MA_baseline β_N error 5.8 % Q=15, P_fus=2538 MW
SPARC_V2C β_N error 22.2 % Q=15, P_fus=840 MW
Interferometer phase RMSE 0.003379 rad 3 channels
Neutron rate relative error 3.0 %
Thomson voltage RMSE 6.11e-07 V 3 channels

Heating & Neutronics

Metric Value Unit Notes
Best Q (ITER-like scan) 15.00 — Target: Q ≥ 10
Q ≥ 10 achieved Yes — 0.80 × 10²⁰ m⁻³
P_aux at best Q 10.0 MW Auxiliary heating
P_fus at best Q 1564.0 MW Fusion power
T at best Q 24.8 keV Ion temperature
ECRH absorption efficiency 99.0 % 170 GHz, 1st harmonic, 20 MW
Tritium Breeding Ratio (total) 1.1409 — 3-group, 80 cm, 90% ⁶Li
TBR fast group 0.0278 — 14.1 MeV neutrons
TBR epithermal group 0.2257 — Slowed neutrons
TBR thermal group 0.8875 — Thermalized

Disruption & Control

Metric Value Unit Notes
Disruption prevention rate (SNN) 100.0 % 50-run ensemble
Mean halo current peak 1.408 MA
P95 halo current peak 2.111 MA
Mean RE current peak 0.014 MA
P95 RE current peak 0.021 MA
ITER halo+RE contract pass (stress lane) Yes — Requires prevention>=90%, P95 halo<=3.4 MA, P95 RE<=1.0 MA
HIL control-loop P50 latency 24.5 μs 1000 iterations
HIL control-loop P95 latency 37.8 μs
HIL control-loop P99 latency 96.5 μs
Sub-ms achieved Yes — Total loop: 31.9 μs

Real-Shot Validation

Metric Value Unit Notes
Disruption recall 1.00 — 6/6 disruptions detected
Disruption FPR 0.00 — 0/10 false alarms
Disruption detection Yes — recall ≥ 0.6 and FPR ≤ 0.4
Transport tau_E RMSE 0.0969 s 53 shots
Transport within 2σ 72 % Gate ≥ 80%
Transport validation Yes —
Equilibrium ψ pass fraction 67 % 12/18 files
Equilibrium q95 pass fraction 100 % 18/18 files
Equilibrium validation Yes —
Data provenance Mixed — Real SPARC/ITPA + template-generated DIII-D disruption shots
Overall real-shot pass Yes —

Disturbance Rejection

Controller Scenario ISE Settling (s) Overshoot Stable
PID VDE 1.08e-05 0.0843 0.0200 Yes
H-infinity VDE 4.86e-05 0.0894 0.0397 Yes
MPC VDE 5.27e-06 0.0553 0.0171 Yes
SNN VDE 7.74e-02 1.9999 0.3100 Yes
PID Density ramp 5.96e-05 3.9999 0.0091 Yes
H-infinity Density ramp 9.08e-04 3.9999 0.0316 Yes
MPC Density ramp 2.06e-05 3.9999 0.0048 Yes
SNN Density ramp 1.55e-01 3.9999 0.3143 Yes
PID ELM pacing 7.18e-07 2.9999 0.0013 Yes
H-infinity ELM pacing 1.37e-05 2.9999 0.0045 Yes
MPC ELM pacing 4.86e-07 2.9593 0.0011 Yes
SNN ELM pacing 1.15e-01 2.9999 0.3104 Yes

Manufactured-Source Equilibrium Parity (Solov'ev lane)

Case ψ NRMSE q NRMSE Axis error (m) Passes
ITER-like 0.000 0.000 0.000 Yes
SPARC-like 0.000 0.000 0.000 Yes
Spherical-tokamak 0.000 0.676 0.000 Yes
KSTAR-like 0.000 0.000 0.000 Yes
SPARC-high-kappa 0.000 0.000 0.000 Yes

Provenance: synthetic manufactured-source parity lane (release default when FreeGS backend is unavailable).

Overall ψ NRMSE: 0.000 (threshold: 0.11). Overall: PASS

Disruption Transfer-Generalization

Group Shots Disruptions Safe Recall FPR
Source 5 1 4 1.000 0.000
Target 11 5 6 1.000 0.000

Transfer efficiency (target/source recall): 1.000 | Overall: PASS

Disruption Threshold Optimization

Metric Value Notes
Optimal bias -5.0
Optimal threshold 0.99
Recall 1.00 6/6
FPR 0.50 5/10
Pareto score 0.60 recall − FPR
Shots evaluated 16 6 disruptions, 10 safe

Legacy Surrogates

Metric Value Unit Notes
Neural transport MLP surrogate tau_E RMSE 0.0607 s ITPA H-mode confinement time
Neural transport MLP surrogate tau_E RMSE % 13.5 % 20 samples

Validation Summary

Lane Status Key metric
QLKNN Transport PASS test_rel_l2 = 0.0943
Real-shot validation (mixed real+template) PASS recall=100%, FPR=0%
Confinement ITPA RUN RMSE = 0.0969 s
3D Force Balance RUN reduction = 3.5×
Q ≥ 10 PASS Q = 15.0
TBR > 1.05 PASS TBR = 1.1409
ECRH absorption RUN 99.0%
Disruption detection PASS recall=100%
HIL sub-ms PASS P50 = 24.5 μs
Solov'ev manufactured-source parity PASS ψ NRMSE = 0.000
Transfer generalization PASS eff=1.000, target_recall=1.000

Documentation & Hero Notebooks

Official performance demonstrations and tutorial paths:

  • examples/neuro_symbolic_control_demo_v2.ipynb (Golden Base v2)
  • examples/platinum_standard_demo_v1.ipynb (Platinum Standard - Project TOKAMAK-MASTER)

Legacy frozen notebooks:

  • examples/neuro_symbolic_control_demo.ipynb (v1)

All benchmarks run on the environment listed above. Artifact-based lanes load pre-computed JSON from artifacts/ and weights/. Timings are wall-clock and may vary between machines. Re-run with python validation/collect_results.py to reproduce.