Simulation of multiphase flow with thermochemical reactions: A review of computational fluid dynamics (CFD) theory to AI integration

This review explores the integration of Computational Fluid Dynamics (CFD) and Artificial Intelligence (AI) in the modeling of multiphase flows and thermochemical systems, which have the characteristics of nonlinear interactions, complex geometries, and high computational costs. These systems, found in diverse applications such as chemical reactors, energy production, and environmental modeling, present significant challenges in accurately simulating dynamic fluid behaviors. Traditional CFD approaches, while mathematically rigorous, often struggle with convergence efficiency, mesh sensitivity, and physical boundary constraints in high-dimensional or reactive flow environments. Recent developments in machine learning (ML), particularly deep learning (DL) and physics-informed neural networks (PINNs), have catalyzed a paradigm shift in fluid dynamics modeling. Data-driven models now enable real-time inference, surrogate modeling, and multiscale learning, surpassing the conventional limitations of CFD solvers. These techniques leverage vast datasets, often generated by simulations or experiments, to develop models capable of making accurate predictions without the need for extensive computational resources. Frameworks such as neural operators and hybrid physical-statistical models offer not only improved scalability but also enhanced robustness across diverse flow regimes, from turbulent flows to complex reactive systems. Despite this promise, AI-enhanced CFD still faces key challenges. Many AI models depend heavily on empirical data rather than physics-based simulations, limiting their generalizability and physical consistency. Inverse modeling techniques, such as reinforcement learning, remain in their early stages, reducing their effectiveness for parameter optimization in heat transfer and fluid flow. Additionally, AI models often struggle to generalize across unfamiliar flow regimes—such as transitions from laminar to turbulent or reactive flows—restricting their broader applicability. These challenges highlight the need for more robust and interpretable AI-CFD frameworks. Nonetheless, promising results have been achieved. For instance, PINNs applied to the lid-driven cavity flow problem demonstrated a maximum mean squared error of 7.38 × 10−4 in the horizontal and 5.99 × 10−4 in the vertical direction compared to OpenFOAM solutions. Furthermore, inference cost scales linearly with grid resolution, and computational speed exceeds that of traditional solvers by factors ranging from 12 to 626, showcasing substantial gains in efficiency, scalability, and accuracy. The integration of AI into CFD holds the potential to revolutionize simulation capabilities, opening new frontiers for industrial applications and scientific research involving complex fluid systems.