● AI · Coalbed Methane ANNBP Sorption Isotherm IATMI 2014

Langmuir parameter prediction
from sorption-isotherm data

Neural-network backpropagation trained on sorption isotherm analysis (SIA) data to predict Langmuir volume (V_L) and Langmuir pressure (P_L) for CH₄ and CO₂ in coalbed methane — compressing a six-input network into a three-input predictor without losing accuracy.

R² 0.949
CH₄ pattern recognition
R² 0.961
CO₂ pattern recognition
R² 0.965
3-input prediction network

Why Langmuir parameters matter for CBM

In coalbed methane, gas is stored by adsorption on the coal matrix — not as free gas. The sorption isotherm governs how much gas the coal can hold at a given pressure, and its two defining parameters — Langmuir volume (V_L) and Langmuir pressure (P_L) — feed directly into original gas-in-place and sorption-scale recovery-factor calculations.

Obtaining V_L and P_L traditionally requires sorption isotherm analysis (SIA) on core samples, which is expensive and not always available on every well. The question this study addressed: can a neural network learn the mapping from the rest of the SIA measurements to V_L and P_L — and if so, which inputs actually matter?

  • Inputs: depth, pressure, temperature
  • Inputs: helium-pycnometry density
  • Inputs: ash content, moisture content
  • Outputs: V_L and P_L (CH₄ network)
  • Outputs: V_L and P_L (CO₂ network)
  • Dataset: 73 CH₄ samples, 48 CO₂ samples

Three-stage ANNBP workflow

The study used artificial neural network backpropagation (ANNBP) — a feedforward network with tansig/purelin activations — organised into three sequential stages that each serve a distinct purpose:

  • Pattern recognition — train the full 6-input network on 90% of the data, validate and test on the remaining 10%
  • Performance-driver identification — re-run the network with each input systematically removed, then rank inputs by ΔR² to isolate the true signal carriers
  • Parsimonious prediction — retrain using only the top-three drivers, confirming that the simpler network retains predictive accuracy

The key insight: a 6-input network is not the same as a 6-input model. Three of the six inputs carry most of the predictive signal — the remaining three add complexity without accuracy. The reduced network is cheaper to deploy and easier to interpret.

Three inputs match six — cleanly

Pattern-recognition networks for both gases achieved R² close to unity. When the input set was reduced to the three most influential parameters, prediction quality held — composite R² = 0.965 across CH₄ and CO₂ combined.

Stage Gas Inputs
Pattern recognitionCH₄60.949
Pattern recognitionCO₂60.961
Prediction (reduced)CH₄ + CO₂30.965

The three drivers are thermodynamically sensible

For CH₄ the ranked drivers are density → moisture → temperature; for CO₂ they are moisture → temperature → density. The ordering differs between the two gases, but the set of three is the same — consistent with adsorption physics, where pore-surface availability (density), competition with water for adsorption sites (moisture), and kinetic energy (temperature) are the primary levers on sorption capacity.

A compact predictor for OGIP and RFsc

The reduced three-input network delivers Langmuir parameters accurate enough to feed OGIP and sorption-scale recovery-factor calculations for saturated coal — without requiring the full SIA measurement suite on every sample. The same methodology extends naturally to under-saturated coal as a follow-on study.

This work was published at the IATMI Annual Scientific Conference 2014 and became the methodological precursor to later work predicting Langmuir parameters together with gas content and permeability directly from open-hole logs.

/ Conference paper · IATMI 2014
Rachmat, S. & Hadad, A.F. (2014).
Pattern Recognition to Predict Langmuir Parameters in Coalbed Methane Using Neural Network.
Proceedings of the IATMI Annual Scientific Conference, Jakarta.
Published by Indonesian Association of Petroleum Engineers (IATMI)