Abstract
In this chapter, we introduce the solution space for high-fidelity models based on partial differential equations and the finite element model. The manifold learning approach to model order reduction requires simulated data. Hence, learning projection-based reduced order models (ROM) has two steps: (i) an offline step for the computation of simulated data and for consecutive machine learning tasks, (ii) an online step where the reduced order model is used as a surrogate for the high fidelity model. The offline step generates a train set and a validation set of simulated data. The accuracy and the generalisation of the reduced order model is evaluated in the online step by using a test set of data forecast by the high-fidelity model. The test set aims also to check the computational speedups of the reduced-order model compare to the high-fidelity model.
The Challenge of High-Fidelity Simulations
In engineering fields like structural mechanics and fluid dynamics, High-Fidelity Models (HFM) based on Partial Differential Equations (PDEs) and the Finite Element Method (FEM) are the gold standard for accuracy. However, these models are computationally expensive, often requiring hours or days to solve a single simulation on a complex domain $\Omega$.
This computational cost creates a bottleneck for real-time applications, optimization loops, and digital twins. The goal of Reduced-Order Modeling (ROM) is to create a surrogate model that is fast enough for an "online" phase—running in seconds—while maintaining an acceptable level of accuracy compared to the original HFM.
Methodology: From Parameter Space to Solution Manifold
The core idea behind ROM is that the solutions $u(\mu)$ of a PDE, while living in a very high-dimensional space, often lie on a much lower-dimensional manifold governed by the input parameters $\mu$.
Building these models involves a two-step workflow:
- Offline Phase: This is the data generation and learning stage. Researchers use the HFM to generate "snapshots"—solutions at different parameter values $\mu$ chosen via Design of Experiments (DoE). These snapshots are used to train the reduced space using techniques like Proper Orthogonal Decomposition (POD) or Nonlinear Manifold Learning (e.g., Deep Auto-Encoders).
- Online Phase: The learned ROM is used as a surrogate. Instead of solving the full system of size $N$, the model solves a reduced system of size $n$, where $n \ll N$.
Key Techniques for Efficiency
Hyper-Reduction (DEIM & ECM)
Simply reducing the number of variables is often insufficient for nonlinear problems. Calculating nonlinear internal forces usually requires integrating over the entire original mesh, which keeps the computational cost tied to the full mesh size $N$. Hyper-reduction techniques, such as the Discrete Empirical Interpolation Method (DEIM) and Empirical Cubature Method (ECM), allow the model to perform calculations only on a "Reduced Integration Domain" (RID), truly decoupling the ROM from the original mesh size.
The $n < N^{1/3}$ Rule
A practical guideline for engineers is that for a ROM to be truly competitive with modern sparse finite element solvers, the dimension of the reduced space $n$ should ideally be less than the cube root of the original degrees of freedom $N$. This accounts for the $O(n^3)$ cost of solving the dense reduced system versus the efficient sparse solvers used in HFMs.
Conclusion and Future Directions
While projection-based ROMs offer massive speedups, challenges remain. Geometric variability—where the domain $\Omega$ itself changes, such as in crack propagation—remains a complex frontier. Additionally, many hyper-reduction schemes are "intrusive," requiring deep access to the source code of the FEM solver. Future work focuses on making these methods more robust to geometric changes and developing non-intrusive methods that can work with commercial software packages.
Summary updated