TL;DR
This paper introduces and proves the concept of "verifiability" for data-driven reduced order models (ROMs) in fluid dynamics. They demonstrate that their proposed data-driven variational multiscale ROM (DD-VMS-ROM) is verifiable, meaning that a small error in the ROM's closure model leads to a small error in the overall ROM solution. This is a crucial step towards building more reliable and trustworthy ROMs for complex turbulent flows.
What Problem Does This Paper Address?
Reduced Order Models (ROMs) are powerful tools for accelerating simulations of complex physical systems, especially in fluid dynamics where full-order models (FOMs) can be computationally prohibitive. However, for turbulent flows, ROMs often require "closure models" to account for the effects of unresolved scales. The development of these closure models, particularly data-driven ones, has outpaced their theoretical understanding. A key challenge has been to guarantee that an accurate closure model leads to an accurate ROM solution. This paper addresses this gap by formally defining and proving the "verifiability" of a data-driven ROM closure model, establishing a rigorous link between the accuracy of the closure and the accuracy of the ROM.
Methods
The authors extend the concept of "verifiability" from Large Eddy Simulation (LES) to the ROM context. A ROM closure is deemed verifiable if a small error in the closure model implies a small error in the ROM solution. They then propose and analyze a data-driven variational multiscale ROM (DD-VMS-ROM). The core of their approach involves:
- ROM Spatial Filtering: Using the ROM projection as a spatial filter to decompose the ROM solution into large and small scale components.
- Data-Driven Closure Model: Employing a linear ansatz for the ROM closure term and determining its parameters by solving a least-squares problem using FOM data.
- Physical Constraints: Incorporating a physical constraint (mean dissipativity) into the least-squares problem to improve stability and accuracy.
- Theoretical Proof: Formally proving the verifiability of the DD-VMS-ROM using mathematical analysis, including the use of discrete Gronwall's lemma.
Datasets
The paper uses numerical simulation data generated from:
- 1D Burgers Equation: A standard test case for fluid dynamics models, used to demonstrate the DD-VMS-ROM's verifiability.
- 2D Flow Past a Cylinder: A more complex benchmark problem, used to further validate the theoretical findings for different Reynolds numbers (Re=100 and Re=1000).
The "truth" solution for these simulations is obtained by projecting FOM data onto the ROM space, effectively using the FOM as the ground truth.
Key Insights and Ideas
- Formalizing Verifiability for ROMs: The paper provides the first formal definition and theoretical investigation of "verifiability" for ROM closure models. This is a significant step towards building trust in data-driven ROMs.
- Bridging Data-Driven Models and Theory: The DD-VMS-ROM is shown to be verifiable, demonstrating that the data-driven approach, when properly constrained and analyzed, can lead to theoretically sound models.
- The Importance of Closure Accuracy: The verifiability proof establishes a direct relationship: a more accurate ROM closure model (smaller closure error) leads to a more accurate ROM solution (smaller ROM error).
- Physical Constraints Enhance Reliability: The inclusion of a mean dissipativity constraint is shown to be crucial for proving the verifiability of the DD-VMS-ROM.
Key Results
The paper presents both theoretical and numerical results:
- Theoretical Proof of Verifiability: Theorem 5.2 formally proves that the DD-VMS-ROM, under certain conditions (small time step), is verifiable.
- Numerical Validation:
- For the 1D Burgers equation, numerical experiments show that as the closure error (η(L²)) decreases, the ROM error (E(L²)) also decreases, consistent with the verifiability concept. The relationship often exhibits a slope greater than or equal to 1, indicating good performance.
- For the 2D flow past a cylinder at Re=100 and Re=1000, similar trends are observed. The DD-VMS-ROM demonstrates that reducing the closure error leads to a reduction in the ROM error, supporting the theoretical findings. The linear regression slopes generally align with or exceed the theoretical prediction.
Limitations and Future Work
- Limitations:
- The theoretical proof of verifiability relies on a "small time step" assumption, which is common in numerical analysis but might limit its direct applicability to very large time steps.
- The verifiability definition and proof are primarily in the L² norm.
- The numerical investigation for the flow past a cylinder at Re=100 showed a slope slightly below 1 for a low-dimensional ROM (r=4), which the authors attribute to insufficient data points for accurate slope estimation.
- Future Work:
- Investigating the verifiability of other types of ROM closure models (functional, structural, or different data-driven approaches).
- Extending the verifiability concept to ROMs trained on experimental data.
- Exploring different norms (e.g., H¹ norm) for the least-squares problem and verifiability definition.
- Adapting concepts from classical LES verifiability to the ROM setting.
Summary updated