Location:Home >> Research >> Research Progress

"Dongfang" Supercomputer Empowers Intelligent Design of Scramjet Engines

Date: Aug 01, 2025

Quantitative analysis of complex combustion process is critical in the design and optimization of propulsion systems for hypersonic vehicles. Traditional approaches rely heavily on manual interpretation of massive flow field data, which entails high computational costs, low efficiency in pattern recognition, and difficulties in conducting cross-condition correlation analyses—factors that significantly hinder the design optimization process.

Recently, the research team led by Guihua Shan at the Computer Network Information Center of the Chinese Academy of Sciences (CAS), in collaboration with Wei Yao’s team at the Institute of Mechanics, CAS, developed an intelligent system for clustering and summarizing combustion time-series flow fields of scramjet engines. This achievement was made possible through the domestic high-performance computing capabilities of the "Dongfang" supercomputing platform and provides a new analytical tool for studying supersonic combustion dynamics.

The research involved pattern recognition on 210 high-resolution large eddy simulation (LES) cases, each with 18.48 million grid cells, covering a wide range of engine initial conditions—static pressure from 0.8 to 2.1 MPa, static temperature from 565 to 830 K, and water vapor concentration from 7.8% to 14%. Supported by the "Dongfang" supercomputing platform, the teams have conducted high-throughput numerical simulations and data analysis totaling 6 TB. By applying pretrained Vision Transformer (ViT) models for spatially weighted embeddings of key flow field variables such as pressure and OH fields, and combining these with dimensionality reduction and clustering algorithms, the system achieved automatic recognition and trajectory tracking of combustion pattern evolutions. To address the issue of noisy global feature clustering, the system improves clustering performance and enhances scientific separability of phenomena—achieving at least 60% reduction in computational resource consumption compared to traditional methods when reaching steady combustion states.

An integrated "computation-analysis-interpretation" framework was built: under the "Dongfang" supercomputing environment, the team employed the Dynamic Zoning Flame Model (DZFM) proposed by the Institute of Mechanics, CAS, to efficiently complete 210 high-fidelity simulations of scramjet engine scenarios (which would have taken several years using conventional computing methods). A temporal trajectory similarity metric was designed to analyze the evolution of combustion across cases with different initial conditions. Leveraging a knowledge base of expert annotated flows of cluster centers, the system uses a Vision-Language Model (VLM) to automatically generate descriptions and summaries for single flow frame and entire simulation process—dramatically reducing expert annotation workload from over 3,000 images to just 16 cluster centers. This enhances the VLM’s ability to extract and describe scientific phenomena, effectively supporting comprehensive scientific analysis. Case studies demonstrate improvements in both the efficiency of combustion phenomenon interpretation and the accuracy of identifying convergence of steady combustion temporal trajectories.

The related work, titled TemporalFlowViz: Parameter-Aware Visual Analytics for Interpreting Scramjet Combustion Evolution, has been accepted by ChinaVis 2025, a premier domestic conference in the field, and recommended for publication in the Journal of Visualization, a leading journal in visual analytics. The project is supported by the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant No. XDB0500301), undertaken by the Computer Network Information Center, CAS.


Appendix: