A High-Fidelity Digital Twin Framework for Proppant Transport: Integrating GPU-Accelerated Barracuda Virtual Reactor MP-PIC with Graph Neural Networks

Abstract

Abstract Optimizing proppant distribution across multiple clusters is essential for maximizing reservoir contact, yet Heel-to-Toe bias and limited entry inefficiencies remain pervasive. While high-fidelity Computational Fluid Dynamics (CFD) is an industry standard for visualizing these unseen downhole dynamics, a bridge is required between high-resolution simulation and real-time operational decision making. This work presents a workflow where GPU-accelerated Barracuda Virtual Reactor Multiphase Particle-in-Cell (MP-PIC) simulations provide the essential synthetic data to train a Graph Neural Network (GNN)-based Reduced Order Model (ROM), enabling rapid predictive capabilities for completion design optimization. The workflow leverages the MP-PIC solver to model dense slurry flows in 3D multifracture domains, accounting for full Particle Size Distribution (PSD) and non-Newtonian fluid drag effects. Using the MP-PIC solver's massive GPU parallelization, high-fidelity training data is generated in days, varying cluster spacing (5 to 20 m) and pumping rates (10 to 20 bpm) to capture complex phenomena like stress shadowing and momentum-driven particle bypassing. These Eulerian fields train a GNN-ROM architecture featuring spatial graph convolutions and Fourier feature embeddings for explicit time-conditioning. Because the GNN-ROM takes the inlet flow rate as an input parameter, it can be driven by a measured surface pumping curve to predict placement during a treatment, forming the core of a framework that, coupled with field diagnostics of flow allocation, extends toward a real-time digital twin. The GNN-ROM enables instantaneous, single-shot predictions of proppant distribution at any point in the treatment schedule, providing a step-change in speed compared to conventional physics-based simulators. For the single-fracture case, the model predicts the experimentally measured bed-height profiles with mean absolute errors ranging from 0.0064 to 0.0150 m. For the unseen 15 bpm multifracture case, it closely reproduces the CFD-predicted heel-to-toe proppant distribution, achieving a spatial correlation coefficient of 0.957 and a range-normalized RMSE of 11.2%. Individual full-field predictions require approximately 1.5-2 s on average, while complete transient sequences are generated 44-74 times faster than the corresponding CFD simulations. The ROM model accurately predicts proppant bed morphology and cluster-level starvation, replicating the high-fidelity physics of the MP-PIC solver at the speed required for field-scale decision support. This work demonstrates that high-fidelity CFD can serve as the data engine for fast, accurate surrogate models of proppant transport. By coupling MP-PIC simulations with a GNN-based ROM, completion design questions such as the influence of limited entry and cluster spacing can be explored in seconds rather than hours to days, providing a physics-grounded alternative to purely empirical design workflows.