Generative ML Inversion for Borehole Sensing

The patent applies preliminary action by pre-training a generative machine learning model on a large dataset of forward simulation results before actual inversion is needed. This pre-computed knowledge is stored in the model's parameters, enabling fast inversion without performing expensive simulations during the actual inversion process. The model learns the inverse mapping from measurements to material properties in advance, so when real measurements are input, the inversion can be performed rapidly using the pre-acquired knowledge.

The patent uses an intermediary approach by introducing a generative machine learning model that is specifically trained to perform the inverse mapping from measurements to material properties. This intermediary model acts as a bridge between the forward simulation capabilities and the inversion task, learning the complex inverse relationship during training. The model stabilizes the inversion process by providing a consistent, data-driven mapping that accounts for the uncertainties and non-linearities in the measurement-to-properties relationship, rather than attempting direct inversion of the forward model.

The patent applies the copying principle by creating a surrogate model (generative machine learning model) that replicates the inverse mapping functionality of traditional inversion methods. Instead of performing computationally intensive traditional inversion calculations each time, the system uses the pre-trained neural network model that has copied the essential inversion knowledge. This surrogate model produces results that closely match traditional methods but with significantly reduced computational energy requirements, as the heavy lifting was done during the one-time training phase.