3D Digital Twin

What Is a Digital Twin?

A digital twin is a virtual replica of a physical system that runs in real time. For a PEM electrolyzer, it means:

• Your trained PINN predicts per-cell voltages at each operating point

• A Lattice Boltzmann CFD solver models water/gas flow in the channels

• A Three.js viewer renders the 4-cell stack in 3D with live colour-coded health

The twin you just launched runs your best_12param.pt checkpoint — the model you trained in Parts 1–6.

---

Architecture

results/best_12param.pt
        │  MODEL_PATH env var
        ▼
digital_twin/backend/server.py   ← FastAPI + LBM CFD at 30 Hz
        │  WebSocket JSON frames
        ▼
digital_twin/digital_twin_3d.html  ← Three.js 3D stack
        │  sensor readings: state.cells[]
        ▼
Predict panel  ← independent ML fault classifier
        │
        ▼
Fix panel  ← corrective action recommendation

One SSH tunnel port (8000) serves both the WebSocket and the HTML viewer.

---

Running the Twin

From your demonstrator notebook, run the Part 7 cells:

# 1. Point backend at your trained model
MODEL_PATH = str(Path('results/best_12param.pt').resolve())

# 2. Start backend (reads MODEL_PATH from environment)
proc = subprocess.Popen(
    ['python', 'digital_twin/backend/server.py'],
    env={**os.environ, 'MODEL_PATH': MODEL_PATH},
)

# 3. Open http://localhost:8000

Remote VM? Forward both ports in your SSH command:

ssh -L 8888:localhost:8888 -L 8000:localhost:8000 user@vm.sigma2.no

---

The 4 Independent Panels

Panel	Purpose	What it reads
🔵 TWIN	Live stack health from PINN + CFD	PINN voltages, LBM temperature
🧪 SIMULATE	Inject 8 failure scenarios	User button click
🔮 PREDICT	ML fault classification	state.cells[] sensors only
🔧 FIX	Corrective action recommendation	Predict output

Key design: the PREDICT panel reads only sensor values — it never reads the scenario button you clicked. This mirrors how a real fault-detection system must operate: it only sees measurements, never the ground truth failure label.

Your model, detecting a fault it was never told about. The PREDICT panel raises ⚠️ “Cell 3 membrane wearing out” from live sensor values — without ever reading which scenario button was clicked. Cell 4 health has fallen to 82%; the FIX panel is ready to act.

---

Connecting to Your Model

The backend uses the MODEL_PATH environment variable:

PINN_MODEL_PATH = os.environ.get(
    'MODEL_PATH',
    str(Path(__file__).parent.parent.parent / 'demonstrators' / 'student_model.pt')
)

The notebook cell sets MODEL_PATH=results/best_12param.pt — your freshly trained checkpoint — without any backend code changes.

---

Exercises

Exercise 1 — Fault detection from sensors alone

1. Select the Technician role (gear icon → Technician)

2. Click Thermal Runaway in the SIMULATE panel

3. Watch the PREDICT panel — does it detect “thermal” or “runaway” from sensor data alone?

4. Note: PREDICT never sees which scenario button you clicked

Exercise 2 — Apply Fix and observe trajectory

1. Start Membrane Dry-out

2. Wait for PREDICT to recommend a fix

3. Click ⚡ APPLY FIX in the FIX panel

4. Watch the health chart: solid lines (actual) bend upward; dashed lines project where health would have gone without the fix

Exercise 3 — Export telemetry to pandas

1. Run a scenario for 30 seconds

2. Click Export Replay (top bar) — downloads replay.json

3. Load it back into a notebook cell:

import json, pandas as pd
with open('replay.json') as f:
    frames = json.load(f)
df = pd.DataFrame([
    {'tick': fr['tick'], 'cell': ci, 'health': c['membraneHealth']}
    for fr in frames
    for ci, c in enumerate(fr['state']['cells'])
])
df.pivot(index='tick', columns='cell', values='health').plot()