Comparative Study on Neural Operators for Reservoir Simulation

  • Homepage
  • Event
  • Comparative Study on Neural Operators for Reservoir Simulation
Dr. Waleed Diab

Dr. Waleed Diab

Solutions Architect for Energy NVIDIA
Wednesday, 28 October 2026 1:30 – 2:15 PM
Building 19, Hall 1
Keynote
05 Session Talk 05 Sustainable Environment and Supply of Essential Needs
Presentation

Abstract

Comparative Study on Neural Operators for Reservoir Simulation

Accurate and efficient simulation of subsurface flow dynamics is critical for carbon sequestration monitoring, hydrocarbon recovery optimization, and reservoir management. Traditional numerical simulators, while physically rigorous, often incur prohibitive computational costs that limit their applicability in uncertainty quantification, real-time decision-making, and large-scale optimization workflows. Neural operators have emerged as promising surrogate models capable of learning solution operators directly from simulation data, offering orders-of-magnitude speedups while maintaining accuracy. However, the relative performance of different neural operator architectures across varying reservoir complexities, dimensionalities, and training paradigms remains insufficiently characterized.

This study presents a systematic comparative evaluation of three distinct classes of neural operators for reservoir simulation: Fourier Neural Operator (FNO) variants, Deep Operator Networks (DeepONet) variants, and graph-based architectures represented by XMeshGraphNet. We assess these models on two benchmark datasets spanning different physical phenomena: a 2D open-source CO₂ sequestration dataset, featuring temporal evolution of pressure and saturation fields over 24 time steps, and the Norne field dataset, a full three-dimensional reservoir model with multi-phase flow dynamics. Within the FNO family, we examine standard FNO, U-Net enhanced FNO (U-FNO), and Conv-FNO architectures. For the DeepONet framework, we systematically evaluate two fundamental architectural paradigms: single-branch and multi-branch configurations. Single-branch DeepONets encode the input function through a unified branch network, while multi-branch variants (MIONet) employ separate encoders to process distinct input modalities before fusing their representations. Within each paradigm, we examine multiple branch implementations including standard MLP, U-Net-enhanced, spectral convolution (Fourier) and hybrid branches. The XMeshGraphNet provides a complementary perspective through its mesh-agnostic, message-passing formulation suitable for irregular geometries.

We investigate two training regimes: full spatio-temporal mapping, where models predict the complete spatiotemporal solution in a single forward pass, and autoregressive rollout, where predictions are generated sequentially and fed back as inputs. Furthermore, we examine the role of physics-informed hybrid losses that combine data-driven objectives with physical constraints, including spatial derivative regularization, mass conservation enforcement, and black oil equation residuals. Importantly, our study focuses exclusively on hybrid data-physics approaches rather than purely physics-driven training, reflecting practical deployment scenarios where simulation data is available but physical consistency remains desirable.

Our methodology shows the AI research community how to leverage PhysicsNeMo to accelerate AI surrogate model experimentation and development. Our findings provide guidance for practitioners selecting neural operator architectures based on dataset characteristics, accuracy requirements, and computational constraints. Moreover, the entire code base and datasets utilized in this study are fully open-sourced through the PhysicsNeMo library.

About the speaker

Biography

Waleed Diab is a Solutions Architect for Energy at NVIDIA, where he works with energy companies and ISVs across the Middle East and Europe on GPU-accelerated simulation and AI. His focus spans machine learning surrogates for reservoir simulation, physics-informed and neural operator methods, agentic AI for engineering workflows, and AI infrastructure for the energy sector.

Waleed holds a PhD in Petroleum Engineering from Khalifa University and is a member of SPE, SIAM and IAMG.

Attend this session

The Saudi HPC/AI Conference runs 26–28 October 2026 at KAUST. Attendance is free, but registration is required.