| Abstract Submission No. | ABS-4.1-0017 |
| Title of Abstract | Physics Informed Machine Learning for CO2 plume monitoring |
| Authors | Soumya Singh * |
| Organisation | IIT ISM DHANBAD |
| Address | Diamond hostel, IIT ISM Dhanbad Dhanbad, Jharkhand, India Pincode: 826004 E-mail: 23mc0095@iitism.ac.in |
| Country | India |
| Presentation | Poster |
| Abstract | Monitoring CO₂ storage requires mapping time-lapse seismic changes to subsurface saturation. This work uses time-lapse seismic differences paired with reservoir simulations from the Sleipner benchmark to train a two-stage model: a supervised 3D U-Net teacher that learns from reservoir fields (Vp, Vs, density, porosity) and a student that takes only seismic difference fields and porosity. In this framework, the student model is trained not only to replicate the teacher¿s predictions via knowledge distillation but also to satisfy physics-based constraints. Penalty terms enforcing mass conservation and flow continuity are included in the loss function, ensuring predictions remain consistent with reservoir flow behavior. This integration balances data-driven learning with physical realism, leading to reliable outputs for CO₂ storage monitoring. On the study dataset the student recovers the main plume shape and the injected mass with low voxelwise error; visual slices comparing ground truth, teacher, student, and absolute error show good agreement in the plume core and larger differences at the edges. Remaining uncertainty is largest where seismic contrast is low, porosity or registration errors exist, or seismic noise is high; adding physics constraints reduces these errors but does not remove them entirely. Overall, the approach turns time-lapse seismic signals into fast, physically guided 3D plume estimates suitable for routine monitoring and assessment. |
| Keywords | Keywords: CO₂ sequestration, plume monitoring, physics-informed neural network (PINN), teacher¿student knowledge distillation. |