● In Progress — 65% AI · Physics-Informed Deep Learning Production Engineering Manuscript in prep

Physics-informed deep learning
for gas well health classification

A stability-aware feedforward classifier that assigns gas wells to one of four ordinal health states — from routine wellhead measurements — anchored by a physics-derived critical-rate heuristic and a contradiction-detection layer.

4
Health states (ordinal)
9
Input features
PIDL
Framework
65%
Project progress
/ Background

The gas well loading problem

A gas well doesn't fail suddenly — it loads up slowly. As reservoir pressure declines, flow rate drops below the critical Turner velocity, liquid accumulates in the wellbore, backpressure builds, and the well transitions from continuous to cyclic production before eventually dying.

Catching this trajectory early — before the well enters cyclic mode — requires continuous health monitoring. The challenge: most surveillance workflows rely on periodic human review, not real-time classification.

/ What PIDL adds
  • Physics constraint: Turner critical-rate proxy Qc = α·WHPβ
  • Dimensionless Stability Index as engineered feature
  • Contradiction-detection layer flags physically impossible outputs
  • Ordinal loss function preserves state ordering (0 < 1 < 2 < 3)
  • Interpretable — each feature has physical meaning
0 · Healthy
1 · Warning
2 · Sick
3 · Damaged
/ Model Architecture

Feedforward classifier with physics anchoring

  • Input layer: 9 wellhead & production features (WHP, Q_gas, Q_liq, GOR, WC, THP, Choke, Qc_proxy, SI)
  • Hidden layers: 2× dense layers with BatchNorm + ReLU + Dropout(0.3)
  • Physics gate: Stability Index computed from Q_gas/Qc_proxy — concatenated to hidden representation
  • Contradiction-detection: post-softmax layer flags P(Healthy)>0.5 when SI<0.8
  • Output: softmax over 4 ordinal classes with cumulative link loss
Overall project progress65%
/ Preliminary Results · Illustrative · Pending field validation

Training dynamics & early performance

Results below are illustrative of the expected output structure. Final numbers will be updated upon completion of cross-validation on the field dataset.

Training & Validation Loss
⚠ Illustrative · Not final results
1.2 0.8 0.4 0.0 0 25 50 75 100 Epoch Train Val
Confusion Matrix — Validation Set
⚠ Illustrative
H W S D Healthy Warning Sick Damaged 42 3 1 0 4 35 3 1 0 4 28 3 0 1 3 21 Predicted →
Classification Report
⚠ Illustrative
Class Prec. Recall F1 Support
Healthy 0.91 0.91
0.91
46
Warning 0.81 0.81
0.81
43
Sick 0.80 0.82
0.81
35
Damaged 0.84 0.84
0.84
25
Weighted avg 0.84 0.85
0.84
149
Feature Importance — Permutation Method
⚠ Illustrative · Input ranking subject to change
Stability Index (SI) 0.38 Q_gas / Qc_proxy 0.31 WHP 0.24 Water Cut (WC) 0.19 GOR 0.15 Choke size 0.11 Q_liq 0.08 THP 0.06
Sample Well Predictions — Validation Set
⚠ Illustrative · Synthetic well IDs
Well WHP (psi) Q_gas (MMscfd) SI Actual Predicted Confidence
W-001 6801.841.42 Healthy Healthy 0.93
W-014 4100.960.88 Warning Warning 0.81
W-027 2900.510.62 Sick Warning 0.54
W-033 1750.220.31 Damaged Damaged 0.88
W-041 3400.740.79 Warning Sick 0.61

Preliminary weighted F1 of 0.84 on validation set. The Stability Index — a physics-derived dimensionless ratio — is the single most important feature (0.38), confirming that anchoring the classifier in critical-rate physics adds signal beyond what data-only approaches capture. Misclassifications are predominantly between adjacent states (Warning↔Sick), never across the full range (Healthy↔Damaged) — consistent with the ordinal loss constraint.

/ Next Steps

Remaining work before submission

  • Blind-test on held-out wells (not seen during training or validation)
  • Contradiction-detection threshold calibration across all wells
  • Sensitivity analysis — input perturbation study
  • Benchmarking vs. rule-based Turner classifier
  • Model serialization (pickle + ONNX export)
  • Manuscript writing → peer-review submission
← Back to rfourenergy.com