NS-ROM Project: Cross-Exploration of Model Order Reduction and Machine Learning

NS-ROM Project is an exam project for the Model Order Reduction (MOR) and Machine Learning course, exploring how to combine model reduction techniques with machine learning methods for simulation and prediction of complex systems.

Project original author/maintainer: tommasopantano01; Source platform: GitHub; Original title: ns-rom-project; Original link: https://github.com/tommasopantano01/ns-rom-project; Published date: June 12, 2026.

In scientific and engineering computing, simulation of high-dimensional complex systems faces extremely high computational costs:

• Scenarios: Computational Fluid Dynamics (CFD), structural mechanics, climate models, circuit simulation, etc.;
• Problems: A single simulation takes hours/days, which is unbearable for real-time applications (e.g., model predictive control); parameter scanning/optimization requires thousands of simulations; embedded deployment has limited resources.

The goal of Model Order Reduction (MOR) is to significantly reduce computational complexity while maintaining acceptable accuracy.

Projection-Based Methods

Core idea: Project the high-dimensional state space onto a low-dimensional subspace. Mathematical principle: For the original system dx/dt = f(x, u) (x ∈ Rⁿ), the reduced-order system dxᵣ/dt = Vᵀ * f(V*xᵣ, u) is obtained via projection matrix V ∈ Rⁿˣʳ (r << n). Common techniques:

• POD (Proper Orthogonal Decomposition): Data-based principal component analysis to extract important modes;
• Krylov subspace methods: Match moments of transfer functions, suitable for linear systems;
• Balance truncation: Based on controllability and observability Gramians, retain important states.

Challenges for Nonlinear Systems

• Hyper-reduction: Nonlinear terms after projection still require high-dimensional computation, requiring techniques like DEIM (Discrete Empirical Interpolation Method) and GNAT (Greedy Algorithm for Nonlinear Approximation);
• Parameter dependence: System behavior changes with parameters, requiring parameterized reduced-order models;
• Stability guarantee: Reduced-order models may lose the stability of the original system.

Data-Driven Reduction

Traditional methods require mathematical models of the system; machine learning can learn from data:

• Autoencoders: Neural networks learn low-dimensional representations;
• Manifold learning: Discover the intrinsic low-dimensional structure of high-dimensional data;
• Gaussian processes: Learn error correction for reduced-order models.

Surrogate Models

Use ML to directly replace complex simulations:

• Neural networks: Train networks to approximate input-output mappings;
• Gaussian process regression: Provide prediction uncertainty estimates;
• Random Forest/XGBoost: Non-parametric regression methods.

Physics-Informed Neural Networks (PINN)

Embed physical constraints into training: The loss function includes residuals of physical equations, and outputs satisfy physical laws, suitable for physical consistency constraints of reduced-order models.

Operator Learning

Learn operators from function spaces to function spaces:

• DeepONet: Learn nonlinear operators;
• Fourier Neural Operator: Learn physical operators in the frequency domain; Suitable for solution operators of parameterized partial differential equations.

The 'NS' in the project name refers to Navier-Stokes equations, which describe the basic equations of fluid motion.

Challenges of Navier-Stokes Equations

• Nonlinear convection term: (u·∇)u leads to nonlinearity;
• Multi-scale characteristics: Turbulence covers vortex structures of multiple orders of magnitude;
• High computational cost: Direct Numerical Simulation (DNS) requires extremely high resolution;
• Parameter sensitivity: Changes in parameters like Reynolds number may lead to drastic changes in flow patterns.

Application Scenarios of ROM in CFD

• Real-time simulation: Real-time aerodynamic analysis of aircraft;
• Optimization design: Fast evaluation of wing shape optimization;
• Uncertainty quantification: Analysis of parameter uncertainty propagation;
• Digital twin: Real-time virtual mirror of physical systems;
• Control applications: Fast simulation in model predictive control.

Typical Methods

• POD-Galerkin: Project onto POD modes, derive reduced-order equations via Galerkin projection;
• POD-NN: Use neural networks to approximate reduced-order dynamics;
• Physics-constrained ROM: Maintain mass and momentum conservation during reduction.

Based on the project name and description, possible technical routes include:

Scenario 1: Pure Data-Driven Surrogate Model

1. Collect high-fidelity simulation data (flow fields under different parameters);
2. Extract low-dimensional features using POD or autoencoders;
3. Train neural networks to learn reduced-order dynamics;
4. Use neural networks to predict flow fields under new parameters.

Scenario 2: Physics-Constrained Reduced-Order Model

1. Start from Navier-Stokes equations, perform Galerkin projection onto POD modes;
2. Handle hyper-reduction of nonlinear terms (DEIM);
3. Use machine learning to correct errors of reduced-order models;
4. Solve the reduced-order system quickly in the online phase.

Scenario 3: Parameterized Reduced-Order Model

1. Sample in the parameter space to generate training data;
2. Learn the mapping from parameters to reduced-order coefficients;
3. Generate reduced-order models quickly for new parameters;
4. Combine with Gaussian processes for uncertainty quantification.

Academic Value

Model order reduction is an important direction in computational science and engineering, involving:

• Mathematical foundations: Linear algebra, functional analysis, differential equations;
• Computational methods: Numerical analysis, high-performance computing;
• Machine learning: Deep learning, Gaussian processes, optimization theory;
• Application fields: Fluid mechanics, structural mechanics, heat transfer, electromagnetism.

Industrial Applications

• Aerospace: Aircraft aerodynamic optimization, real-time simulation;
• Automotive: Engine combustion simulation, aerodynamic design;
• Energy: Nuclear reactor safety analysis, wind farm optimization;
• Biomedical: Blood flow simulation, organ modeling.

Cutting-Edge Trends

• Non-intrusive reduction: No need to modify original simulation code;
• Adaptive reduction: Dynamically adjust reduction dimension based on error;
• Multi-physics coupling: Handle multi-physics phenomena like fluid-structure-heat;
• Uncertainty quantification: Reduced-order models assist fast Monte Carlo simulations.

The NS-ROM project represents cutting-edge exploration in the intersection of model order reduction and machine learning. By combining traditional model reduction techniques with modern machine learning methods, it demonstrates how to balance high-fidelity simulation accuracy and real-time computational efficiency.

For learners, this project provides an excellent opportunity to understand complex system modeling, reduction theory, and the application of machine learning in scientific computing. It is a comprehensive training in physical understanding, mathematical modeling, and computational implementation.

With the growing demand for applications like digital twins and real-time simulation, the importance of model order reduction technology is becoming increasingly prominent. Mastering knowledge in this field will lay a solid foundation for research and work in directions such as scientific computing, engineering simulation, and AI for Science.