8e62716ce4
Core changes: - New phase-state features (PHASE_STATE_KEYS, ILLUSION_PHASE_KEYS) with obs dynamics - Derivative and absolute output modes (output_mode="deriv"|"absolute") - predict_v23_deriv() with integration support for closed-loop - Offline multi-step rollout evaluator (eval_rollout.py) Key results: - Illusion 0.75L/1L: phase-state+error-state+abs achieves 0.974/0.958 closed-loop with zero action history features — proving the new route works - Karman re100: phase-state+abs reaches 0.699 (vs 0.901 with action history) - 1.5L confirmed as bang-bang regime (R2=0.12 for linear SINDy) - Feature ablation: 6-dim phase-state outperforms 16-dim full-lag in closed-loop Documentation: - docs/SR_analysis_results.md: comprehensive analysis report - docs/HANDOVER_SR_ANALYSIS.md: handover notes for next coder - 6 figures in docs/figures/SR_analysis/ - Updated README.md, sindy_sr_notes.md, sindy_sr_knowledge.md - Updated configs.py with generalization scenes Co-authored-by: Cursor <cursoragent@cursor.com>
226 lines
11 KiB
Markdown
226 lines
11 KiB
Markdown
# SR Analysis: Phase-State SINDy Results Report
|
|
|
|
> Date: 2026-06-15
|
|
> Project: DynamisLab — Active hydrodynamic cloaking and illusion using DRL on a fluidic pinball.
|
|
> Analysis pipeline: SINDy (STLSQ) + feature engineering for interpretable control law extraction.
|
|
|
|
---
|
|
|
|
## Executive Summary
|
|
|
|
Three key findings from the SR analysis pipeline:
|
|
|
|
1. **Illusion 0.75L + 1L: New route validated.** Phase-state features + error-state + absolute action output achieves closed-loop similarity of **0.96+** (97% of PPO) with **zero action history features**. This proves that physically meaningful control laws can be extracted without relying on action memory.
|
|
|
|
2. **Illusion 1.5L: Regime shift identified.** The 1.5L target exhibits bang-bang/saturated control (alpha range [-8, 8], autocorrelation r=0.07). Linear SINDy is fundamentally inadequate (R2=0.12). This is a regime boundary, not a modelling failure.
|
|
|
|
3. **Karman: State representation still incomplete.** The same phase-state approach reaches 0.699 (vs 0.901 baseline with action history). The problem is not the output form (derivative vs absolute) but insufficient input state information for the Karman scene.
|
|
|
|
---
|
|
|
|
## 1. Methodology
|
|
|
|
### 1.1 Feature Architecture (Final)
|
|
|
|
Three feature levels were tested for fitting `alpha = f(state)` or `d(alpha)/dt = g(state)`:
|
|
|
|
| Level | Features | Dim | Description |
|
|
|-------|----------|:---:|-------------|
|
|
| Static | u_m, u_a, u_c, v_a, Cd_tot, Cd_rear, Cl_tot, Cl_diff | 8 | Current-step physics only, no memory |
|
|
| Phase-state | u_a, du_a/dt, Cl_tot, dCl_tot/dt, Cd_tot, Cd_rear (+error terms) | 6+4 | Oscillation phase + rate + drag feedback |
|
|
| Full-lag | Static + lag-1 of all 8 | 16 | Brute-force temporal context (baseline) |
|
|
|
|
### 1.2 Output Modes
|
|
|
|
Two output targets were compared:
|
|
|
|
- **Derivative mode**: predict d(alpha)/dt, then integrate `alpha(t) = alpha(t-1) + dt_c * dalpha/dt`
|
|
- **Absolute mode**: predict alpha directly, no integration needed
|
|
|
|
### 1.3 Evaluation Metrics
|
|
|
|
Models are evaluated on:
|
|
1. **One-step R2**: fit quality on training data
|
|
2. **Offline multi-step rollout**: 1/5/20/50 step recursive prediction on held-out PPO data
|
|
3. **CFD closed-loop**: full CFD environment with SINDy control law, DTW similarity vs target
|
|
|
|
---
|
|
|
|
## 2. Illusion Results
|
|
|
|
### 2.1 Three-Scene Summary
|
|
|
|
| Scene | S | Old v23 (a_lag) | New phase+error+abs | PPO baseline | % of PPO | Features | Action history? |
|
|
|------|:--:|:----------------:|:-------------------:|:------------:|:--------:|----------|:---------------:|
|
|
| 0.75L | 400 | 0.908 | **0.974** | 0.972 | 100.2% | 10-dim (ILLUSION_PHASE) | **No** |
|
|
| 1L | 600 | 0.962 | **0.958** | 0.973 | 98.5% | 10-dim (ILLUSION_PHASE) | **No** |
|
|
| 1.5L | 800 | 0.926 | **N/A** | 0.945 | — | — | Bang-bang regime |
|
|
|
|
**Key insight**: The 0.75L scene's new route **outperforms** the old v23 (0.974 vs 0.908), while 1L matches it within 1.5%. This definitively proves that "action history is not necessary."
|
|
|
|
### 2.2 Illusion 1L Front Coefficients (Phase-State, Absolute Output)
|
|
|
|
```
|
|
alpha_F = f(u_a, du_a/dt, Cl_tot, dCl_tot/dt, Cd_tot, Cd_rear, Cd_err, Cl_err, dCd_err/dt, dCl_err/dt)
|
|
R2 = 0.987
|
|
```
|
|
|
|
Key contributors (sorted by |coef|): Cd_rear > Cd_tot > Cd_err > Cl_tot > Cl_err > dCl_err/dt > dCd_err/dt
|
|
|
|
The formula is dominated by **drag-based feedback** (Cd_tot, Cd_rear, Cd_err), with the oscillation phase (u_a, Cl_tot) providing secondary modulation.
|
|
|
|
### 2.3 Illusion 1.5L — Regime Shift Evidence
|
|
|
|
| Metric | 1L | 1.5L | Ratio |
|
|
|--------|:--:|:----:|:-----:|
|
|
| Alpha range (front) | [-0.36, 1.11] | **[-8.0, 8.0]** | 10x |
|
|
| 1-step autocorrelation | 0.957 | **0.065** | — |
|
|
| d(alpha)/dt std | 0.12 | **6.07** | 50x |
|
|
| Linear SINDy R2 | 0.987 | **0.124** | — |
|
|
|
|
The 1.5L controller operates at saturation limits (maximum rotation speed), flipping between extremes. This is consistent with the frequency-doubling control strategy reported in the confirmation report.
|
|
|
|
---
|
|
|
|
## 3. Karman Results
|
|
|
|
### 3.1 Ablation Study (re100)
|
|
|
|
| Configuration | Feat | Output | R2 | Closed-loop | % of PPO |
|
|
|--------------|:----:|:-----:|:--:|:----------:|:--------:|
|
|
| old v23 (a_lag1 dominant) | 14+3 | alpha | 0.996 | **0.901** | 94.4% |
|
|
| **Phase-state -> abs (best new)** | **6** | **alpha** | **0.965** | **0.699** | 73.3% |
|
|
| Phase-state -> deriv | 6 | dalpha/dt | 0.837 | 0.656 | 68.8% |
|
|
| Phase-state + mu -> abs | 9 | alpha | 0.979 | 0.700 | 73.4% |
|
|
| Expanded 10-dim -> abs | 10 | alpha | 0.980 | **0.580** | 60.8% |
|
|
| Full-lag -> deriv | 16 | dalpha/dt | 0.939 | 0.619 | 64.9% |
|
|
| Static -> deriv | 8 | dalpha/dt | 0.321 | 0.745 | 78.1% |
|
|
|
|
**Key insight from ablation**:
|
|
|
|
1. **Phase-state + absolute output is the best new route** (0.699), but still well below the action-history baseline.
|
|
2. Adding extra static features (expanded 10-dim) **hurts** closed-loop despite higher R2 — classic covariate shift.
|
|
3. The static->deriv paradox: low R2 (0.321) but good closed-loop (0.745), because no-memory models are naturally robust to rollout divergence.
|
|
4. Mu modulation doesn't help at single-Re; its value will appear in cross-Re fitting.
|
|
|
|
### 3.2 Cross-Re Generalization (old v23 model)
|
|
|
|
| Re | Type | Closed-loop | Note |
|
|
|:--:|:----:|:----------:|------|
|
|
| 50 | Training | 0.582 | Low-frequency shedding |
|
|
| 100 | Training | **0.901** | Default / best |
|
|
| 200 | Training | 0.793 | Moderate degradation |
|
|
| 400 | Training | 0.664 | High-Re challenge |
|
|
| 25 | Unseen (subcritical) | 0.567 | Below Hopf bifurcation |
|
|
| 70 | Unseen | 0.577 | Between Re50 and Re100 |
|
|
| 150 | Unseen | 0.595 | Between Re100 and Re200 |
|
|
| 300 | Unseen | 0.541 | Outer extrapolation |
|
|
|
|
### 3.3 One-Step R2 vs Closed-Loop Paradox
|
|
|
|
A key phenomenon discovered during this work: **high one-step R2 does not predict good closed-loop performance** when temporal features (lags, derivatives) are present in the input. This is because:
|
|
|
|
- **Training**: features use ground-truth PPO observations
|
|
- **Deployment**: features use SINDY-controlled observations (distribution shift)
|
|
- Temporal features amplify this shift recursively
|
|
|
|
---
|
|
|
|
## 4. Roadmap to Next Steps
|
|
|
|
### 4.1 Illusion: Proceed to PySR
|
|
|
|
The 0.75L and 1L scenes are ready for symbolic regression. Recommended input:
|
|
|
|
```python
|
|
ILLUSION_PHASE_KEYS = [
|
|
"u_a", "du_a_dt", # oscillation phase
|
|
"Cl_tot", "dCl_tot_dt", # lift dynamics
|
|
"Cd_tot", "Cd_rear", # drag feedback
|
|
"Cd_err", "Cl_err", # force error
|
|
"dCd_err_dt", "dCl_err_dt", # error dynamics
|
|
]
|
|
```
|
|
|
|
Output: absolute alpha (non-dimensional action), no integration.
|
|
|
|
First do separate PySR for each scene, then compare formula structure for shared backbone.
|
|
|
|
### 4.2 Karman: CCD/OID for state completion
|
|
|
|
The 0.699 ceiling suggests missing state variables. Candidate directions:
|
|
- Recirculation zone length / reattachment point
|
|
- Wake centerline deflection
|
|
- POD mode coefficients for phase completion
|
|
- CCD modes correlated with control action
|
|
|
|
### 4.3 SR / CCD / OID Integration Framework
|
|
|
|
```
|
|
┌─────────────────────────────────────┐
|
|
│ Control Objective │
|
|
│ (stealth / illusion / erase) │
|
|
└──────────┬──────────────────────────┘
|
|
│
|
|
┌──────────▼──────────┐
|
|
│ DRL Policy │
|
|
│ (PPO + Sin act) │
|
|
└──────────┬──────────┘
|
|
│
|
|
┌────────────────┼────────────────┐
|
|
│ │ │
|
|
┌────────▼──────┐ ┌─────▼──────┐ ┌──────▼─────────┐
|
|
│ SINDy / SR │ │ CCD / OID │ │ Validation │
|
|
│ obs -> act │ │ obs -> z │ │ CFD closed- │
|
|
│ white-box │ │ structure │ │ loop + DTW │
|
|
│ control law │ │ analysis │ │ │
|
|
└────────┬──────┘ └─────┬──────┘ └──────┬─────────┘
|
|
│ │ │
|
|
└────────────────┼─────────────────┘
|
|
│
|
|
┌──────────▼──────────┐
|
|
│ Interpretable │
|
|
│ Control Mechanics │
|
|
│ obs -> z -> act │
|
|
│ -> structure -> │
|
|
│ -> signature │
|
|
└─────────────────────┘
|
|
```
|
|
|
|
---
|
|
|
|
## 5. Code Changes Summary
|
|
|
|
### Files Modified (8 core files)
|
|
|
|
| File | Changes |
|
|
|------|---------|
|
|
| `src/SR_analysis/utils/feature_builder.py` | PHYSICS_FEAT_KEYS, ILLUSION_ERR_KEYS, PHASE_STATE_KEYS, ILLUSION_PHASE_KEYS, KARMAN_EXPANDED_KEYS; obs dynamics (lag1 + derivative); error-state computation; 1D target_forces support |
|
|
| `src/SR_analysis/utils/sindy_fitter.py` | compute_action_deriv(); get_feature_matrix_deriv() with output_mode="deriv"|"absolute"; include_mu support |
|
|
| `src/SR_analysis/sindy/run_all_v2.py` | run_single_scene_deriv(); run_joint_karman_deriv(); --deriv, --phase, --karman-expand, --karman-mu, --output-mode, --augment-level CLI |
|
|
| `src/SR_analysis/validate/run_closed_loop.py` | predict_v23_deriv() with output_mode; mode="abs" branch; load_sindy_coefs returns "mode" |
|
|
| `src/SR_analysis/validate/run_closed_loop_illusion.py` | Support predict_v23_deriv; auto-detect SINDy mode from coefs |
|
|
| `src/SR_analysis/utils/__init__.py` | Export new constants and functions |
|
|
| `src/SR_analysis/sindy/wrap_joint.py` | Parameterized for karman/illusion |
|
|
| `src/SR_analysis/validate/eval_rollout.py` | **New**: offline multi-step rollout evaluation |
|
|
|
|
### Key Design Principles
|
|
|
|
1. **No action history in features** for all new routes (PHASE_STATE_KEYS, ILLUSION_PHASE_KEYS)
|
|
2. **Time-normalized derivatives** for cross-scene compatibility: `dx/dt = (x(t) - x(t-1)) / dt_c`
|
|
3. **Error-state** for Illusion: Cd_err, Cl_err encode current-to-target deviation
|
|
4. **Output flexibility**: both derivative and absolute modes supported
|
|
|
|
---
|
|
|
|
## 6. Figures
|
|
|
|
All figures in `docs/figures/SR_analysis/`:
|
|
|
|
- **fig1_illusion_comparison.png**: Illusion 3-scene comparison bar chart
|
|
- **fig2_karman_ablation.png**: Karman re100 ablation across 7 configurations
|
|
- **fig3_karman_generalization.png**: Karman cross-Re generalization (training vs unseen)
|
|
- **fig4_r2_vs_closedloop.png**: One-step R2 vs closed-loop paradox scatter
|
|
- **fig5_illusion_coefficients.png**: Illusion 1L front feature coefficients
|
|
- **fig6_roadmap.png**: Research progress roadmap
|