Real-time True Resistivity Estimation for LWD Tools

The patent introduces an intermediary processing layer between the raw EM measurements and the final resistivity interpretation. A neural network model acts as a mediator that takes the apparent resistivity measurements and the estimated boundary distance as inputs, then outputs a corrected true formation resistivity value that compensates for the polarization horn effects. This intermediary processing step transforms the problematic raw measurements into accurate formation properties.

The patent changes the parameter space by not only measuring apparent resistivity but also estimating the distance to the boundary layer. By adding this additional parameter (boundary distance), the system can distinguish between measurements affected by boundary effects and those that are not, enabling correction of the polarization horn effects through the neural network model that processes both parameters together.

The patent replaces complex mechanical or iterative computational inversion systems with a neural network model. Instead of using traditional iterative inversion algorithms that require multiple calculations and assumptions about formation geometry, the system uses a pre-trained neural network that directly maps measurements to corrected resistivity values. This substitution maintains reliability while managing complexity through the use of trained artificial intelligence rather than complex real-time computational mechanics.

The patent performs preliminary action by pre-training the neural network model before field deployment. The neural network is trained offline using synthetic data that covers a wide range of formation scenarios, boundary distances, and resistivity values. This preliminary training allows the system to handle complex boundary effect corrections during actual logging operations without requiring complex real-time computations, thereby improving reliability while keeping the operational system relatively simple.

The patent performs the computationally intensive minimization of differences between expected and actual measurement values in advance, during the offline training phase of the neural network. The neural network is trained on synthetic data where the optimal parameters have already been determined through iterative optimization. This preliminary calculation allows the system to achieve high measurement precision during actual logging operations without performing real-time iterative optimization, thus eliminating the time loss while maintaining accuracy.

The patent creates a simplified copy or approximation of the complex iterative optimization process through the neural network model. Instead of performing the full minimization calculation in real-time, the system uses the neural network's learned parameters and weights that replicate the optimization results. This copying approach allows rapid inference during logging operations while maintaining the precision benefits of the original optimization method.