Two-Stage NMR T2 Mapping for Pore Throat Size Prediction

The patent replaces traditional mechanical/mathematical transformation methods with a machine learning-based approach. A neural network model is trained to map NMR relaxation time distributions to pore throat size distributions, capturing non-linear and heterogeneous relationships that traditional linear or empirical methods cannot adequately represent. This substitution of transformation methodology significantly improves measurement precision while managing complexity through data-driven learning rather than complex analytical formulations.

The patent transforms the problem by changing the approach from direct mathematical transformation to a two-stage process: first transforming NMR relaxation time to NMR pore size using a neural network, then transforming NMR pore size to MICP pore throat size using another neural network. This parameter transformation strategy allows each stage to handle specific aspects of the relationship, improving overall accuracy by breaking down the complex non-linear transformation into manageable segments with dedicated models optimized for their specific tasks.

The patent creates synthetic training data that replicates the characteristics of real core analysis data. By generating artificial datasets that mimic the relationships between NMR measurements, porosity, and pore throat size, the model can be trained on large volumes of data without requiring equivalent amounts of expensive and time-consuming real core analyses. This copying approach allows the model to learn from virtual representations of real-world data, significantly reducing the time and cost burden while maintaining prediction accuracy.

The patent performs preliminary transformations and preprocessing of data before training the neural network models. By pre-processing NMR relaxation time data, converting it to NMR pore size distributions, and preparing the training datasets in advance, the system reduces the computational and temporal requirements during the actual training phase. This preliminary action strategy allows for more efficient use of training data and reduces the overall time needed to develop accurate prediction models.

The patent segments the transformation process into two distinct stages, each handled by a separate neural network model. The first network transforms NMR relaxation time to NMR pore size, while the second network transforms NMR pore size to MICP pore throat size. This segmentation allows each model to be optimized for its specific transformation task, improving overall accuracy by allowing specialized learning for each stage rather than attempting to capture the entire complex non-linear relationship in a single model.

The patent introduces an intermediate dimension (NMR pore size) between the input (NMR relaxation time) and the final output (MICP pore throat size). By transforming through this intermediate representation, the model can capture the complex non-linear relationships more effectively. This dimensional intermediate step allows the system to break down the complex transformation into manageable stages, improving measurement precision while keeping each individual network relatively simple.