Detailed benchmark results across eight scientific domains. Same cognitive architecture — rigorously evaluated with traceable, reproducible evidence.
Drift-wave turbulence governs cross-field particle and energy transport in tokamak edge plasmas. We benchmark UCogNet's cognitive controller against classical baselines (PID, MPC) and three 2026 neural-operator surrogates on the Hasegawa-Wakatani 2D model — the standard testbed for resistive drift-wave dynamics.
🔬
128 × 128
L = 40
α = 0.1
κ = 0.5
ν = 0.005
D = 0.005
dt = 0.005
1 000 + 200 warmup
📐
⚡
Energy stability
🌊
Flux reduction
🌀
Enstrophy control
📊
Spectral fidelity
🔄
Zonal flow fraction
⏱️
Correlation time
🔒
Confinement
🏆
| # | Controller | Composite v3 | Spectral fid. |
|---|---|---|---|
| 🥇 | FNO Surrogate[Li'21] | 0.7187 | 0.922 |
| 🥈 | UCogNet Legacy | 0.7253 | 0.927 |
| 🥉 | NeuOp-Transformer2026[vdW'26] | 0.7293 | 0.835 |
| 4. | FI-Conv2026[Chen'26] | 0.7347 | 0.867 |
| 5. | UCogNet Enhanced | 0.7435 | 0.844 |
| 6. | UCogNet Ω2026 | 0.7490 | 0.819 |
| 7. | PID | 0.7585 | 0.822 |
| 8. | MPC | 0.7645 | 0.789 |
● UCogNet Legacy ranks 2nd on this single run (seed = 42) with highest spectral fidelity (0.927). ✱ FNO, FI-Conv, NeuOp-Transformer are analytical surrogate implementations — not trained model checkpoints from those papers. See multi-seed table for statistical summary.
🎲
| Controller | Mean ± 95% CI | Wins |
|---|---|---|
| UCogNet EnhancedBEST | 0.7219 ± 0.015 | 3/6 |
| NeuOp-Transformer‡ | 0.7285 ± 0.005 | 0/6 |
| UCogNet Legacy | 0.7311 ± 0.022 | 2/6 |
| UCogNet Ω | 0.7400 ± 0.016 | 1/6 |
● UCogNet Enhanced: lowest observed mean (0.7219 ± 0.015). CIs overlap for all pairs; no pairwise difference is statistically significant (e.g. Enhanced vs Legacy: t = −0.779, p = 0.471, n = 6). Win count (3/6) is descriptive.