Consistency and prior falsification of training data in seismic deep learning: Application to offshore deltaic reservoir characterization

Deep learning (DL) applications of seismic reservoir characterization often require the generation synthetic data to augment available sparse labeled data. An approach for generating training consists specifying probability distributions modeling prior geologic uncertainty on properties and forward A falsification is critical establish consistency distribution with real With help a case study facies classification convolutional neural networks (CNNs) from an offshore deltaic reservoir, we have highlighted several practical nuances associated DL models We highlight issue overfitting CNNs propose regularization strategies address it. demonstrate efficacy our proposed by CNN making robust predictions 3D partial stack