Dispersion curve inversion is a key step in obtaining the acoustic velocity of formations around wells and information evaluation. Traditional dispersion inversion methods, based on wellbore sound propagation theory, suffer from low efficiency and multimodality issues. Methods based on deep learning rely on extensive synthetic data for training, with the high economic costs and resource investment limiting their widespread application. To further improve the efficiency of dispersion inversion and reduce algorithm development costs, we have developed a lightweight, deep learning-based dispersion inversion workflow. Firstly, we use parameters with high sensitivity to create a small-scale dataset to train a neural network with a residual structure. By imposing monotonicity constraints on the model output, we achieve rapid and effective simulation of dispersion using less than 5% of the data size required by existing methods. Secondly, we decouple the target dispersion labeling and dispersion fitting to enhance data preprocessing efficiency. Utilizing the neural network’s batch processing capabilities, we employ a grid search method with increasing resolution, achieving inversion results in just three iterations. Lastly, we convert the trained neural network model into an open standard format and deploy it into the onsite data processing workflow. Combined with effective quality control, this approach yields reliable dispersion inversion results.

This content is only available via PDF.
You can access this article if you purchase or spend a download.