Hybrid Neural PDE and Conditional GAN Framework for Sparse Co₂ Plume Prediction.

Digital twin architectures for geological carbon storage demand uncertainty-aware surrogates capable of rapid plume forecasting under extreme data scarcity. Traditional physics-based simulators are computationally expensive; pure data-driven models lack principled uncertainty quantification. This work presents a hybrid Neural Posterior Density Estimation–Conditional GAN (NPDE-CCGAN) framework that integrates physics-aware loss weighting, symmetry-based data augmentation, and amortized Bayesian inference to enable robust CO₂ saturation prediction on 99.70% zero-inflated monitoring data. Applied to Sleipner, the framework achieves R² = +0.007269 across nine test layers, a 297-point improvement from baseline collapse (R² = -297.0), and demonstrates generalization via per-sample R² ranging +0.0070 to +0.0149. The approach is computationally efficient (12 hours CPU training, seconds per inference) and enables Monte Carlo uncertainty quantification for real-time CCS risk assessment.