Speaker
Description
We present a physics-aware deep learning framework for predicting heat flow in heterogeneous geothermal reservoirs. The proposed approach integrates graph neural networks (GNNs) with Transformer-based temporal modeling to serve as a fast and accurate surrogate for conventional reservoir simulators. Spatial representations are constructed through coefficient-aware algebraic multigrid (AMG) coarsening, enabling physics-informed tokenization of heterogeneous permeability and porosity fields on graphs. Temporal evolution is modeled in a latent space using a Transformer architecture, allowing uniform long-term time-step prediction under realistic operational conditions. A dataset of two-dimensional synthetic geothermal reservoir simulations is generated using the MATLAB Reservoir Simulation Toolbox (MRST), incorporating incompressible fluid flow and coupled conductive–advective heat transport in thermal doublet configurations with varying well placements. The proposed model is trained and evaluated against high-fidelity numerical simulation results. The results demonstrate that the GNN–Transformer framework accurately predicts thermal behaviour while achieving substantial reductions in computational cost compared to traditional simulators. These findings highlight the potential of deep learning surrogates for efficient geothermal reservoir forecasting, management, and optimization.
| Country | United Kingdom |
|---|---|
| Student Awards | I would like to submit this presentation into the Earth Energy Science (EES) and Capillarity Student Poster Awards. |
| Acceptance of the Terms & Conditions | Click here to agree |
