top of page
Adolfo Puime Pires - Confirmed Speaker - MRPM 2026

Adolfo Puime Pires

A Two-Stage Galerkin Physics-Informed Neural Network Approach for Forward and Inverse Nuclear Magnetic Resonance Modeling at the Pore Scale

Nuclear Magnetic Resonance (NMR) is widely used in reservoir characterization to evaluate pore-scale properties and fluid behavior in porous media. However, solving forward and inverse pore-scale NMR problems remains computationally challenging due to the coupled diffusion-relaxation mechanisms governing magnetization dynamics. Conventional numerical methods often require fine spatial discretization and high computational cost, particularly in inverse workflows. This work proposes a Pre-Trained Galerkin Physics-Informed Neural Network (pG-PINN) (Constante et al., 2026) approach for solving direct and inverse pore-scale NMR problems in spherical pores. The methodology combines Galerkin-based domain decomposition with physics-informed learning to approximate the Bloch–Torrey diffusion-relaxation equation under diffusion-controlled relaxation conditions. Multiple compact neural subnetworks act as adaptive basis functions, improving training stability and solution accuracy near pore boundaries, where relaxation effects dominate. For the forward problem, the model predicts the spatiotemporal evolution of magnetization and reconstructs NMR relaxation responses for different pore radius, diffusion coefficients, and surface relaxivities. For the inverse problem, the pG-PINN estimates surface relaxivity directly from synthetic NMR relaxation responses. A two-stage pre-training strategy is introduced to accelerate convergence and improve parameter identification. The proposed approach was validated against analytical solutions for spherical pore geometries. Results demonstrate that the pG-PINN achieves high accuracy with reduced computational cost compared with conventional numerical approaches while maintaining robust inverse estimation capability under noisy conditions. The methodology provides an efficient and flexible alternative for AI-driven pore-scale NMR modeling and digital rock physics applications.

​

1) Luis Constante, Adolfo Pires, Martín A. Díaz-Viera. Galerkin physics-informed neural network approach for inverse problems. Application to estimation of tracer dispersion in porous media. Computational Geosciences. Springer. 2026.

bottom of page