sankalpsthakur/forge-pump-surrogate-multiruntime

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.