Forge Pump Surrogate — Multi-runtime
A 1,478-parameter, clean-room pump surrogate exported from one PyTorch MLP to four deployable formats:
| Runtime | Artifact |
|---|---|
| ONNX Runtime / ONNX Runtime Web | onnx/model.onnx |
| PyTorch | pytorch/model.ts, pytorch/model_state.pt |
| TensorFlow | tensorflow/model.keras |
| LiteRT / TensorFlow Lite | tensorflow/model.tflite |
The architecture is a 6 → 32 → 32 → 6 ReLU MLP with input/output normalization embedded in the exported graph. It was trained only on forge-pump-digital-twin-synthetic.
Inputs and outputs
Input order:
speed_fraction, static_head_m, system_k, voltage_fraction,
ambient_temp_c, inlet_pressure_bar
Output order:
flow_m3h, total_head_m, input_power_kw, winding_temp_c,
npsh_margin_m, efficiency_fraction
Test-set metrics
These scores measure approximation of a deterministic synthetic generator—not field accuracy.
| Target | MAE | RMSE | R² |
|---|---|---|---|
| Flow (m³/h) | 0.08441 | 0.11749 | 0.99903 |
| Total head (m) | 0.59088 | 0.83321 | 0.99914 |
| Input power (kW) | 0.03165 | 0.05005 | 0.99814 |
| Winding temperature (°C) | 0.26721 | 0.40810 | 0.99917 |
| NPSH margin (m) | 0.07605 | 0.10571 | 0.99977 |
| Efficiency fraction | 0.00361 | 0.00701 | 0.99751 |
Cross-runtime gate
Thirty-two frozen test vectors passed an absolute tolerance of 1e-3:
| Comparison against PyTorch | Maximum absolute difference |
|---|---|
| ONNX Runtime | 1.5259e-5 |
| TensorFlow | 1.5259e-5 |
| LiteRT/TFLite | 1.5259e-5 |
Exact receipts are in conformance_receipt.json, test_vectors.json, and metrics.json.
ONNX CPU example
import numpy as np
import onnxruntime as ort
session = ort.InferenceSession("onnx/model.onnx", providers=["CPUExecutionProvider"])
x = np.array([[0.80, 35.0, 0.25, 1.0, 25.0, 1.2]], dtype=np.float32)
y = session.run(None, {"features": x})[0]
print(y)
edge/inference_onnx.py adds named inputs/outputs. edge/tag_map.yml shows generic, read-only examples for OPC UA, Modbus input registers, and Siemens S7 symbolic tags. Siemens is a trademark of Siemens AG; this project is independent and is not endorsed by Siemens.
Measured runtime scope
One CPU-only container measurement produced ONNX batch-1 median 0.0061 ms and p95 0.0067 ms; PyTorch batch-1 median was 0.0335 ms. These are smoke-test receipts, not device latency SLOs. See latency_receipt.json for the environment and repetitions.
CUDA was not available in the validation environment, so this release makes no CUDA performance or kernel claim.
Safety boundary
This repository is read-only advisory software. It contains no network client and no PLC, OPC UA, S7, or Modbus write call. It must not bypass or replace PLC/SIS/ESD logic, permissives, interlocks, hardwired protection, or operator authority. The example thresholds are demonstration values, not equipment setpoints.
Provenance
All data and equations were created clean-room for this public release. No customer telemetry, vendor curves, CAD, BOM, nameplate, plant identifier, private model weight, or proprietary source is included.
Rebuild and verify
The generator, training/export script, runtime adapters, manifests, hashes, history, and tests are included. The published build was verified with PyTorch 2.13.0, ONNX Runtime 1.28.0, TensorFlow 2.21.0, and a CPU-only execution provider.
Interactive browser demo: forge-pump-edge-twin-lab
Measured evaluation review — 2026-10-07
All eight vectors supplied in the public test_vectors.json pass PyTorch/ONNX/saved-output comparison at the original 1e-3 bound. The older conformance receipt reports 32 vectors; this review does not replay the other 24 from a saved public vector file.
Report and retained failures · Whole-profile evaluation scope. No external rank-one position or industrial deployment qualification is established. Original scientific objects and model weights are preserved.