An ANFIS model trained on 35 mercury injection capillary pressure datasets from 10 wells — predicting MICP curves from open-hole logs, blind-tested across two fields, and confirming a major gas reservoir with strong sealing capacity.
Mercury injection capillary pressure data serves three distinct purposes in reservoir evaluation, each involving different parts of the MICP curve:
MICP₂.₅ — the MICP value at 2.5% mercury saturation — was used as the target output, representing the threshold entry pressure most relevant for seal capacity assessment.
Direct MICP measurement requires core samples with laboratory mercury injection testing — equipment and analysis not available for every well.
The six inputs are all constituents of regular open-hole logs: Depth, DT (sonic), GR (gamma ray), RHOB (density), RT (resistivity), and NPHI (neutron).
Data selection, calibration, normalization, and grouping. Six log inputs selected. Resistivity converted to logarithmic scale before normalization. All outputs log-transformed.
ANN (trainlm, 2 hidden layers: 30 & 48 neurons) and ANFIS (cluster radius 0.7) trained on full 6-parameter dataset. Establishes baseline R² for each method.
Networks run without each input parameter; RMSE calculated for each omission. Parameters ranked from highest RMSE drop (most influential) to lowest.
Final networks trained using early stopping (ES) with validation set. Both all-parameter and 3-parameter models built — enabling blind-field testing with minimal inputs.
Best networks selected by R². Prediction network applied to C-8 reservoir — a target with no existing MICP data — to assess sealing capacity.
Sensitivity analysis identified three parameters where omission caused the highest RMSE increase — confirming them as the most influential inputs:
Using only these three parameters (GR, DT, RHOB) reduces the data requirement for new wells while maintaining good prediction accuracy — important for practical deployment.
The blind test result (R² = 0.910) used only GR, DT, and RHOB — the three most influential parameters — applied to a neighbouring field the model had never seen during training. This confirms the model generalizes beyond the training dataset.
After validating the prediction network on the blind-test field, the ANFIS 3-parameter model was applied to the C-8 reservoir — a deeper interval with no existing MICP measurements. GR logs were used to estimate the top and bottom of the C-8 sand.
The MICP predictions across the C-8 interval indicated strong sealing capacity in the overlying cap-rock — confirming C-8 as a significant gas-bearing reservoir with seal integrity sufficient to retain a major hydrocarbon column.