ANN and ANFIS pipelines trained to infer three key coalbed methane reservoir properties — gas content, Langmuir parameters, and permeability — directly from open-hole log data, benchmarked across six training algorithms on 315 field-scale CBM datasets from lignite to bituminous.
CBM differs from conventional reservoirs in a fundamental way: it acts simultaneously as source rock and reservoir rock. Gas is stored in the coal matrix through adsorption, not as free gas in pore space. Three properties govern its behavior:
Direct measurement requires core samples and specialized laboratory analysis — a significant budget item not available for every well.
Open-hole logs, however, are routinely acquired in every well. If a reliable predictive relationship can be established between log responses and these properties, CBM characterization cost drops significantly across an asset.
This research covers CH₄ and CO₂ Langmuir parameters — the two dominant gas compositions in the study area.
Resistivity data converted to log scale before normalization. All inputs normalized to range [−1, 1]. All outputs converted to logarithmic form to avoid negative prediction artifacts. Total: 315 datasets from lignite (young) to bituminous (mature) coal rank.
Network architecture: 4-16-1. Six training algorithms tested, with two generalization techniques (early stopping and Bayesian regularization):
Network: 16-4-1. Cluster radius 0.35. Hybrid optimization combining backpropagation and least squares. Early stopping technique for generalization.
Validation dataset: 4% of total data (training + validation split). Process run to maximum 1,000 iterations. NRMSE used as performance metric throughout.
Process flow: Data prep → Pattern/function identification → Generate prediction networks with generalization → Select best technique per property.
| Property | ANN_GDX | ANN_LM | ANN_RP | ANN_SCG | ANN_BR | ANFIS | Best |
|---|---|---|---|---|---|---|---|
| Gas Content | 0.160 | 0.040 | 0.137 | 0.134 | 0.241 | 0.032 | ANFIS |
| VL CH₄ | 0.165 | 0.127 | 0.158 | 0.164 | 0.169 | 0.106 | ANFIS |
| PL CH₄ | 0.073 | 0.032 | 0.060 | 0.055 | 0.034 | 27.192 | ANN_LM |
| VL CO₂ | 0.220 | 0.150 | 0.172 | 0.235 | 0.196 | 0.462 | ANN_LM |
| PL CO₂ | 0.088 | 0.015 | 0.094 | 0.072 | 0.091 | 0.120 | ANN_LM |
| Permeability | 0.071 | 0.054 | 0.079 | 0.046 | 0.018 | 0.182 | ANN_BR |
Rather than selecting a single best technique, the study recommends the three best techniques per property to capture the range of prediction uncertainty:
Conclusion: Both ANN and ANFIS successfully identify the function between open-hole log data and the three key CBM properties. ANN_LM, ANFIS, and ANN_BR are the best overall algorithms. Using multiple techniques per property provides a range of uncertainty for CBM reservoir evaluation — a more honest representation of prediction confidence than a single-model output.