Using mixture density networks for uncertainty and prediction in seismic reservoir characterization

Uncertainty quantification has received significant focus in the geophysics community in recent years; extending focus from general simulation based approaches to measure the amount of variation in the system under study, as well as more complex Bayesian modalities which incorporate ideas from Bayesian statistic and many other methods not mentioned here. More recently, there has been work in Bayesian approaches to neural networks, which endow the parameters of neural networks with prior distributions and exploiting the flexibility of neural network architectures to model complex behavior. We present an illustration of a more recently developed style of modeling uncertainty, Mixture Density Networks (MDNs), and present an application of the method to a reservoir characterization problem. Our approach provides good performance in prediction of the reservoir properties with corresponding uncertainties. In addition, the approach permits new types of inferences that can help guide practitioners in their study of the reservoir.