Abstract
This survey reviews a decade of collaborative research focused on applying Reduced Order Modeling (ROM) and manifold learning to complex industrial settings, particularly in aeronautics and structural mechanics. By bridging the gap between deep learning architectures (such as VAEs and GANs) and classical physics-based solvers, the methodologies presented enable significant computational speedups. This work addresses the challenges of nonlinearity, geometric variability, and the "black-box" nature of commercial engineering software, facilitating real-time digital twins and large-scale design optimization.
Introduction: The Challenge of High-Fidelity Simulation
Modern engineering relies heavily on high-fidelity simulations like Finite Element Analysis (FEA) and Computational Fluid Dynamics (CFD). While accurate, these methods are computationally expensive, often requiring days to solve a single iteration. This latency creates a bottleneck for iterative design, uncertainty quantification (UQ), and real-time monitoring. Traditional model reduction techniques, such as Proper Orthogonal Decomposition (POD), frequently fail when confronted with strong nonlinearities (e.g., plasticity or radiation), geometric variations, or the inability to access the internal source code of proprietary industrial solvers.
Methodological Advancements
The surveyed research categorizes several key strategies to overcome these limitations:
Hyper-Reduction and Nonintrusive ROMs
The Empirical Cubature Method (ECM) is employed to achieve hyper-reduction, allowing simulations to run by calculating only a small fraction of integration points. To accommodate industrial workflows, "nonintrusive" frameworks treat commercial solvers (e.g., Ansys) as data generators. This allows for the creation of ROMs without modifying the underlying solver code.
Nonlinear Manifolds via Auto-Encoders
To capture complex physical behaviors that linear projections miss, the authors utilize Variational Auto-Encoders (VAEs). These networks identify the "latent space" of physical systems, which is particularly effective for highly nonlinear fluid flow problems where traditional linear methods struggle to maintain accuracy.
ROM-nets: Dictionary-Based Learning
For scenarios involving high parameter variability, a single global reduced model is often insufficient. "ROM-nets" utilize deep learning as a recommender system to select the most appropriate local reduced model from a pre-computed "dictionary" based on current inputs, such as specific thermal loads or geometric defects.
Physics-Informed Manifold Learning
The research integrates physical principles directly into the learning process. This includes using "dissipative modes" to stabilize turbulent flow simulations and treating mesh geometries as points on a Grassmann manifold. The latter enables "mesh morphing," mapping complex, non-parameterized shapes onto a common support for more effective regression.
Industrial Applications and Datasets
The methodologies were validated across diverse industrial domains:
- Aeronautics: Analysis of high-pressure compressors and injectors using 3D Navier-Stokes and transient thermal solvers.
- Structural Mechanics: Elastoviscoplastic computations for high-pressure turbine blades.
- Material Science: Assessment of "crack criticality" using 3D X-ray tomography of welded joints.
- Synthetic Benchmarks: Testing neural network compression on 2D Karman Vortex streets and wave equations.
Key Results and Impact
The integration of these techniques has yielded substantial improvements in engineering workflows:
- Computational Speedups: A structural mechanics application saw a reduction in computation time from 9.5 days to just 6.5 hours.
- Selection Efficiency: ROM-nets selected optimal models 60 times faster than traditional clustering methods.
- Model Compression: A priori compression of CNNs for wave simulators reduced trainable parameters by 80% without loss of accuracy.
- Predictive Accuracy: Multimodal auto-encoders for defect assessment achieved error rates near 5%, with speedups ranging from 3x to 120x.
Conclusion and Future Directions
While the survey demonstrates the power of combining manifold learning with physics, challenges remain. Nonintrusive methods, while practical, may face accuracy trade-offs when solver implementations are opaque. Data scarcity in high-fidelity physics remains a hurdle, though physics-specific data augmentation offers a partial solution. Future work will focus on further integrating "black-box" regressors with manifold learning to address increasingly complex multi-physics scenarios and topological changes in mesh-based models.
Summary updated