A study on multi-exponential inversion of nuclear magnetic resonance relaxation data using deep learning.

Area of Science:

• Geophysics
• Petroleum Engineering
• Data Science

Background:

• Nuclear magnetic resonance (NMR) is vital for oil industry formation evaluation, providing critical reservoir parameters like pore structure and fluid saturation.
• The inversion of NMR relaxation data is inherently ill-posed, leading to inaccuracies in reservoir analysis.

Purpose of the Study:

• To develop and validate a deep learning method for accurate multi-exponential inversion of NMR relaxation data.
• To improve the precision of formation evaluation parameters derived from NMR measurements.

Main Methods:

• A deep learning approach using a multi-scale convolutional neural network (CNN) with an attention mechanism (ATT-CNN) was developed.
• Simulated NMR data, incorporating noise and signal characteristics, were used to train the neural network.
• The ATT-CNN model was validated against simulated data and compared with traditional regularization methods using rock core NMR measurements.

Main Results:

• The ATT-CNN model demonstrated superior performance in denoising and T2 inversion compared to other methods using simulated data.
• Practical application with rock core NMR measurements confirmed the effectiveness of the ATT-CNN model.
• The proposed deep learning method significantly outperformed the conventional regularization method.

Conclusions:

• Deep learning, specifically the ATT-CNN model, offers a more accurate and robust solution for NMR relaxation data inversion.
• This advancement holds significant potential for improving the accuracy of reservoir engineering and formation evaluation in the oil industry.