Researchers at Amirkabir University of Technology deploy one-dimensional convolutional neural networks (1D-CNN) and deep jointly informed neural networks (DJINN) to predict formation permeability from synthetic mud loss data generated by reservoir simulation. They preprocess drilling parameters including depth, mud properties, and formation characteristics, then train and test both models, achieving R2 above 0.97. This approach uses real-time drilling data to provide accurate permeability estimates for reservoir management.

Key points

• Synthetic dataset of 810 cases generated via Eclipse E100 simulates drilling fluid loss across variable depths, formation types, thicknesses, mud densities and viscosities.
• 1D-CNN model comprises one convolutional layer, flattening, two dropouts (0.2) and two fully connected layers using ELU activation, trained with Adam optimizer.
• DJINN maps decision tree structures into deep neural network topology and initial weights before backpropagation fine-tuning, achieving higher regression accuracy.
• Data preprocessing includes normalization to [0,1] and 80/20 train/test splitting, ensuring balanced input distributions and robust model validation.
• DJINN yields training/test R2 of 0.978/0.972 versus 1D-CNN’s 0.968/0.962, enabling near real-time, non-invasive permeability estimation during drilling.

Why it matters: By harnessing drill-time mud loss measurements and AI, this method enables continuous, non-invasive estimation of formation permeability, reducing reliance on costly core sampling and well testing. The high R2 scores demonstrated by DJINN suggest more accurate reservoir models, improving drilling efficiency and hydrocarbon recovery predictions.