Machine Learning Applications · Industrial ML with physics

Live site: https://juanlaigle-hub.github.io/MachineLearningApplications/ · Versión en español

The live demos are confidential: use Request a demo to write to me.

Two real plant problems solved end to end: data generated with validated physics simulators, phenomenology checks, honest training and validation, model export, and a live demo that runs in the browser.

	Project	Industrial problem	Result	Live demo
	Dynamometer card diagnosis	Oil & gas: classify every downhole card of a sucker-rod pumped well into 7 conditions (fluid pound, gas interference, valve leaks, parted rod, pump tagging bottom)	98.7 % test accuracy, 96.7 % on out-of-range wells (engineer rules: 50 %)	Request a demo
	Juice concentrate soft sensor	Food: estimate every minute the °Brix (refractometer backup), HMF and color of a 5-effect apple juice evaporator	°Brix error 1.34 on an unseen step test (naive: 6.45); HMF R² 0.93 from 120 lab samples	Request a demo

How it works, step by step

For non-specialists, one slide per project (in Spanish):

Engineering decisions

• Physics first. Features come from physics: fluid load and stroke ratio in the pump; boiling-point elevation, mass balance and Arrhenius × residence time in the evaporator. The °Brix model is sign-constrained, so it cannot learn that more steam lowers the °Brix.
• Downhole card, not surface card. The same model trained on the surface card drops to 34 % on unseen wells, because that card depends on each well's rod-string dynamics. The data confirm why the industry diagnoses with the downhole card computed with the wave equation.
• Gray box before black box when labels are scarce. With only 20 lab samples the gray-box HMF model already performs at its best (MAE 0.43 mg/kg), while gradient boosting needs 120 to reach 0.46. It also recovers the kinetic constant (0.162 vs. 0.18 true) and extrapolates correctly with temperature.
• Model selection without looking at the test set. Cross-validation grouped by well or campaign, only among models that pass the physics tests. A Ridge with similar error was rejected because its mass-balance coefficient had the wrong sign.
• Plant step tests. In closed loop the °Brix barely moves and the data cannot teach cause and effect. Solved with manual feed steps, planned so they do not trip the interlocks.
• Checks find bugs. The phenomenology tests caught three bugs in the simulators themselves, fixed before training:
  • a valve leak that increased production;
  • a °Brix that decreased along the evaporator;
  • an energy chain with two solutions.

From data to deployment

Each project README has a 12-step table comparing how it is done in a real plant with what was done here:

1. Problem and KPI.
2. Historian extraction.
3. Labeling or lab data.
4. Plant step tests.
5. Cleaning and alignment.
6. Physics-based features.
7. Split by well or by time.
8. Baselines and model comparison.
9. Physics validation.
10. Export with exact verification.
11. Confidence thresholds and operating fallbacks.
12. Drift monitoring.

Layout

ml_cartas_dinamometricas/       dataset generator, executed notebook, exported model, live page
ml_sensor_virtual_evaporador/   historian and lab generator, executed notebook, model, page
assets/                         previews
index.html                      portfolio landing page (GitHub Pages)

The 3D physics simulators that generate the data live in Simulations3D.

Notes

• The models were trained on simulated data. Before operational use they must be validated and retrained with real field or plant data. Each README details its limitations.
• MIT license.