Robust Digital Twin for Geological Carbon Storage

While recent work has shown that conditional neural networks can be used to effectuate prior-to-posterior mappings during ensemble Kalman filtering, their accuracy relies on the quality of the training ensemble, consisting of the simulated time-advanced nonlinear state and possibly nonlinear observations. Its precision also depends on the generative neural networks' ability to generalize. In this work, we will report progress on how to make our Digital Twin more robust with respect to distribution shifts and modeling errors, e.g. due to erroneous choices for the physics within the Forecast Step. To this end, we will follow a two-pronged approach consisting of augmenting the Forecast step, e.g. by including different models for the rock physics that map the Digital Twin's state to seismic properties, and by adding more physics during the prior-to-posterior mapping. In its current implementation, training of our Digital Twin's neural networks relies on relative low-fidelity amortization of the Forecast Ensemble, which is susceptible to distribution shifts. By conducting non-amortized inference informed by the physics and low-fidelity prior-to-posterior mapping, we propose an approach that is more robust than the current implementation of the Digital Twin.