Major research outcome
- Professor Sang Lee's Lab: A Study on Dimensionality Reduction of Solution Reconstruction Methods for a Four-Point Stencil
- 관리자 |
- 2026-05-04 21:03:43|
- 63
- 2026-05-04 21:03:43|
In finite volume method (FVM) based computational fluid dynamics (CFD) analysis, solution reconstruction is a critical component that directly determines spatial accuracy. However, reconstruction methods are inherently high dimensional functions that map an n-point stencil to left and right interface states, making their development and improvement challenging. This issue is particularly pronounced on unstructured meshes.
Recently, Professor Sang Lee's research group proposed an dimensionality reduction technique utilizing the normalization of flow variables. The key insight is a mathematical proof that, reconstruction method must satisfy invariance under scaling and translation, so that normalizing the input variables does not alter solution quality. Based on this idea, the team demonstrated that any four point stencil reconstruction method can be fully represented by six two dimensional scalar functions. A visualization technique was also developed to merge these six functions into a single contour plot, enabling intuitive inspection of a reconstruction method.
The proposed dimensionality reduction technique was applied to representative reconstruction methods including MUSCL and WENO3, as well as a trained artificial neural network (ANN). Solution reconstruction via interpolation of extracted datasets was validated on four benchmark problems. Results confirmed that when the dataset size exceeds 6×200×200, the impact of accumulated interpolation error on solution quality is negligible. In particular, for the ANN based method, adopting the dataset interpolation approach enabled a dramatic reduction in additional computational cost.
This study presents a novel mathematical framework that enables dimensionality reduction and visualization of solution reconstruction methods, providing an important theoretical and practical foundation for the future development of data driven and machine learning based reconstruction schemes, as well as the improvement of conventional numerical methods.
This research was published in Advances in Engineering Software (2025, Vol. 199, 103804), an international journal in the field of engineering software.
https://doi.org/10.1016/j.advengsoft.2024.103804
Recently, Professor Sang Lee's research group proposed an dimensionality reduction technique utilizing the normalization of flow variables. The key insight is a mathematical proof that, reconstruction method must satisfy invariance under scaling and translation, so that normalizing the input variables does not alter solution quality. Based on this idea, the team demonstrated that any four point stencil reconstruction method can be fully represented by six two dimensional scalar functions. A visualization technique was also developed to merge these six functions into a single contour plot, enabling intuitive inspection of a reconstruction method.
The proposed dimensionality reduction technique was applied to representative reconstruction methods including MUSCL and WENO3, as well as a trained artificial neural network (ANN). Solution reconstruction via interpolation of extracted datasets was validated on four benchmark problems. Results confirmed that when the dataset size exceeds 6×200×200, the impact of accumulated interpolation error on solution quality is negligible. In particular, for the ANN based method, adopting the dataset interpolation approach enabled a dramatic reduction in additional computational cost.
This study presents a novel mathematical framework that enables dimensionality reduction and visualization of solution reconstruction methods, providing an important theoretical and practical foundation for the future development of data driven and machine learning based reconstruction schemes, as well as the improvement of conventional numerical methods.
This research was published in Advances in Engineering Software (2025, Vol. 199, 103804), an international journal in the field of engineering software.
https://doi.org/10.1016/j.advengsoft.2024.103804
| Attach File |
|---|