Network-Enhanced Gaussian Process-Assisted Multi-Objective Parameter Optimization of Hybrid WENO Schemes for the Compressible Euler Equations

In hybrid weighted essentially non-oscillatory (WENO) schemes, the coupled effects of smoothness detection, nonlinear weighting, and dissipation control give rise to complex nonlinear responses to parameter changes. This coupling makes it challenging to improve smooth-region accuracy and shock resolution simultaneously. Moreover, the high cost of computational fluid dynamics (CFD) evaluations limits the efficiency of direct multi-objective searches. To address these challenges, we couple a network-enhanced Gaussian process model (NEGPM) with the non-dominated sorting genetic algorithm II (NSGA-II) to optimize hybrid WENO parameters for the compressible Euler equations. NEGPM uses a shared neural encoder supervised by both objectives to capture nonlinear interactions among continuous and discrete parameters. In the learned latent space, two independent Gaussian processes with Matérn-5/2 kernels and automatic relevance determination predict smooth-region error and numerical shock thickness. A two-stage training strategy combines shared-feature learning with kernel hyperparameter estimation via marginal likelihood, enabling target-specific predictions from limited CFD samples. These predictions guide non-dominated search, with the selected design subsequently verified by CFD. The model achieves log-space R2 values of 0.9957 and 0.9853 for the two objectives, respectively, and corresponding Spearman coefficients of 0.9985 and 0.9935. This accuracy and ranking consistency support its use in multi-objective optimization. On the Euler benchmarks considered, the optimized scheme reduces smooth-region error by 27.91% and numerical shock thickness by 10.78% relative to WENO-Z. Moreover, the framework achieves optimized performance comparable to direct CFD-driven NSGA-II while reducing CFD evaluations from 2000 to 97. Total optimization time decreases from 126 h 34 min to 6 h 18 min. The framework thus offers an efficient, data-driven approach to improving numerical performance in high-order shock-capturing scheme design.

Authors

Institutions

Publication Details

Journal
Mathematics
Published
2026-11-12
DOI
https://doi.org/10.3390/math14224165
Primary Topic
Computational Fluid Dynamics and Aerodynamics
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Network-Enhanced Gaussian Process-Assisted Multi-Objective Parameter Optimization of Hybrid WENO Schemes for the Compressible Euler Equations

Yunqing Li, Weiwen Zhao, Rui Tian, Cong Wang et al.
Mathematics
Computational Fluid Dynamics and Aerodynamics
article

Network-Enhanced Gaussian Process-Assisted Multi-Objective Parameter Optimization of Hybrid WENO Schemes for the Compressible Euler Equations

Yunqing Li, Weiwen Zhao, Rui Tian, Cong Wang, Tianyu Li
article en

Abstract

In hybrid weighted essentially non-oscillatory (WENO) schemes, the coupled effects of smoothness detection, nonlinear weighting, and dissipation control give rise to complex nonlinear responses to parameter changes. This coupling makes it challenging to improve smooth-region accuracy and shock resolution simultaneously. Moreover, the high cost of computational fluid dynamics (CFD) evaluations limits the efficiency of direct multi-objective searches. To address these challenges, we couple a network-enhanced Gaussian process model (NEGPM) with the non-dominated sorting genetic algorithm II (NSGA-II) to optimize hybrid WENO parameters for the compressible Euler equations. NEGPM uses a shared neural encoder supervised by both objectives to capture nonlinear interactions among continuous and discrete parameters. In the learned latent space, two independent Gaussian processes with Matérn-5/2 kernels and automatic relevance determination predict smooth-region error and numerical shock thickness. A two-stage training strategy combines shared-feature learning with kernel hyperparameter estimation via marginal likelihood, enabling target-specific predictions from limited CFD samples. These predictions guide non-dominated search, with the selected design subsequently verified by CFD. The model achieves log-space R2 values of 0.9957 and 0.9853 for the two objectives, respectively, and corresponding Spearman coefficients of 0.9985 and 0.9935. This accuracy and ranking consistency support its use in multi-objective optimization. On the Euler benchmarks considered, the optimized scheme reduces smooth-region error by 27.91% and numerical shock thickness by 10.78% relative to WENO-Z. Moreover, the framework achieves optimized performance comparable to direct CFD-driven NSGA-II while reducing CFD evaluations from 2000 to 97. Total optimization time decreases from 126 h 34 min to 6 h 18 min. The framework thus offers an efficient, data-driven approach to improving numerical performance in high-order shock-capturing scheme design.

MathematicsVol. 14(22)
Shanghai Jiao Tong University (CN), Fudan University (CN), Ruijin Hospital (CN), Shanghai Technical Institute of Electronics & Information (CN)
Computational Fluid Dynamics and Aerodynamics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon