Abstract: Extensive monitoring, along with data assimilation based on these measurements, will be essential in the management of gigaton-scale geological CO2 storage operations. In this talk, I will present two data assimilation frameworks for uncertainty reduction in flow and geomechanical responses. First, I will describe the use of a deep-learning-based surrogate for model-based data assimilation. A recurrent-residual U-Net, combined with a hierarchical MCMC method, is applied to predict pressure and CO2 saturation in 3D models. Next, I will present a complementary framework, data-space inversion (DSI), to treat more challenging cases involving coupled flow-geomechanics with multiple uncertainty sources. An adversarial autoencoder (AAE) parameterization is developed for the dimension reduction of spatio-temporal data. The AAE is used in DSI posterior sampling to enable direct forecasting of quantities of interest. In this case, DSI is successfully applied to faulted, 3D heterogeneous systems to predict pressure, saturation, stress components, and key aquifer properties.