Research progress on intelligent flow field modeling method for aircraft

Intelligent flow field modeling methods, by integrating the strengths of deep learning in feature extraction and dynamic response prediction with architectural innovations in MDO (multidisciplinary design optimization), have emerged as research hotspot for achieving efficient modeling of complex flow systems and enhancing high-dimensional performance. This fusion paradigm not only strengthens the coupling between data and physics but also provides aerodynamics design with computationally efficient and physics-consistent solutions through multi-objective optimization mechanisms. A new approach for the deep integration of data knowledge and physical mechanisms was provided, aiming to inspire interdisciplinary innovations in intelligent flow field modeling in aerospace and other fields.

In recent years, data-driven deep learning models have demonstrated revolutionary breakthroughs in flow field modeling. By leveraging end-to-end nonlinear mapping capabilities, these models transcend the limitations of traditional approaches reliant on manually defined variable sets and empirical closure models, significantly enhancing cross-configuration generalization performance under varying operational conditions. While pixel-based CNNs (convolutional neural networks) have been extensively applied to 2D flow field studies, their inability to capture geometric details severely compromises generalization performance when handling complex 2D boundary conditions or 3D flow fields. This limitation has been progressively addressed through point cloud networks (PointNet) and GNNs (graph neural networks), which effectively encode global geometric features in 3D flow modeling.

Intelligent flow prediction methodologies are evolving from purely data-driven paradigms toward physics-constrained hybrid frameworks. This paradigm shift not only improves prediction accuracy but also enhances model robustness and data efficiency, providing more reliable solutions for complex fluid dynamics challenges. PINNs (physics-informed neural networks) exemplify this trend by incorporating fundamental fluid mechanics laws into neural network loss functions through partial differential equation constraints, thereby strengthening physical consistency. Although showing promise across various physical modeling applications, PINN training still faces challenges from gradient discrepancies. Operator learning methods, particularly represented by DeepONet and FNOs (Fourier neural operators), have emerged as pivotal tools for learning mappings between infinite-dimensional function spaces to replace conventional numerical PDE solvers. These approaches effectively integrate the demonstrated strengths of data-driven models while overcoming input dimensionality constraints, positioning themselves as crucial enablers for future CFD (computational fluid dynamics) innovation.

Physics-constrained modeling significantly reduces dependence on large annotated datasets while improving adherence to physical principles, showing substantial potential for enhancing aircraft design efficiency from subsystem-level optimization to integrated system development. The convergence of deep learning with fundamental physics formulations continues to reshape the landscape of fluid dynamics simulation, offering new pathways for solving previously intractable engineering problems.

Intelligent flow field prediction demonstrates substantial application potential in MDO for aircraft, where efficient flow analysis enables accelerated design iteration cycles and enhanced overall design efficiency. Despite preliminary achievements in this domain, current research outcomes remain constrained by notable limitations in applicability, primarily stemming from challenges in acquiring high-fidelity datasets, effectively representing complex boundary geometries, and establishing robust physics-constrained frameworks. Addressing key technical bottlenecks—including improving shape generalization capability, prediction accuracy, and cross-physical-scenario adaptability—is critical for transitioning flow field modeling from theoretical research to practical engineering applications.

Current investigations must prioritize resolving fundamental challenges such as fidelity preservation in cross-scale parameter transfer and interpretability enhancement of intelligent algorithms. Physics-informed architectures demonstrate potential for reconciling data-driven predictions with fundamental conservation laws, thereby improving extrapolation capability in novel aerodynamic configurations. Continued progress in these areas could substantially advance the integration of intelligent flow prediction into MDO workflows, ultimately enabling more efficient exploration of complex design spaces under multiphysics constraints.