Direct Multi?Modal Inversion of Geophysical Logs Using Deep Learning

Geosteering of wells requires fast interpretation geophysical logs, which is a non-unique inverse problem. Current work presents proof-of-concept approach to multi-modal probabilistic inversion logs using single evaluation an artificial deep neural network (DNN). A mixture density DNN (MDN) trained the "multiple-trajectory-prediction" (MTP) loss functions, avoids mode collapse typical for traditional MDNs, and allows prediction ahead data. The proposed verified on real-time stratigraphic gamma-ray logs. predictor outputs several likely solutions/predictions, providing more accurate realistic solutions than deterministic regression DNN. For these curves, model simultaneously predicts their probabilities, are implicitly learned from training geological stratigraphy predictions probabilities obtained in milliseconds MDN can enable better decisions under uncertainties.