A Physics-Based Loss Function to Constrain Neural Network Inversion of 4D Seismic Data

Research output: Contribution to conferencePaperpeer-review


We present a deep neural network application for 4D seismic inversion to changes in pressure, water and gas saturations. The method is applied to a real field data case, where, as is common in such applications, the data measured at the wells is insufficient for training neural networks, thus, the network is trained on synthetic data. As neural networks can extrapolate beyond the training data, it is possible to reach physically inconsistent results when applying the network to unseen data. We present a methodology to incorporate prior physics information to constrain the neural network relations to physically consistent results. This is done by including physics information into the training loss function, guiding the training process towards models that can describe the training data as well as present physically consistent solutions. Using this method, the network models reach precise results even when trained on general non-ideal synthetic datasets.
Original languageEnglish
Number of pages5
Publication statusPublished - Oct 2021
Event82nd EAGE Annual Conference & Exhibition 2021 - Amsterdam, Netherlands
Duration: 18 Oct 202121 Oct 2021


Conference82nd EAGE Annual Conference & Exhibition 2021


Dive into the research topics of 'A Physics-Based Loss Function to Constrain Neural Network Inversion of 4D Seismic Data'. Together they form a unique fingerprint.

Cite this