Numerical simulation is an important method in supercomputing and AI for Science. Due to the high complexity of the physical parameter space in numerical simulations, traditional post-processing methods are inefficient, requiring multiple iterative steps to explore the parameter space. Current state-of-the-art (SOTA) methods only support a unidirectional process from parameters to visualization results, and obtaining the target parameters still requires numerous trials. Furthermore, the generation of scientific visualization results is constrained by strict conditions, leading to the discrete distribution of effective latent vectors. This results in the failure of existing continuous latent vector editing methods from computer vision generative models, preventing them from being applied to scientific visualization prediction image editing tasks.

To address the above challenges, our team from the Advanced Interactive Technology and Application Development Department designed the ParamsDrag model. We innovatively proposed a latent vector jumping-editing method based on continuous input space, successfully overcoming the editing bottleneck caused by the discrete distribution of effective latent vectors. Additionally, the team proposed and designed a visualization generation model based on precise constraints, pioneering a new exploratory approach of "simulation parameters — visualization prediction — physical parameter inversion." Through multiple experiments, it was demonstrated that the ParamsDrag model performs well in visualization prediction and parameter exploration in real scientific numerical simulations, outperforming state-of-the-art (SOTA) methods. This achievement provides strong support for the efficient exploration of the physical parameter space in numerical simulations. The related paper has been accepted by the IEEE Visualization conference (IEEE Vis 2024, a CCF-A conference). The first author of the paper is Associate Professor Guan Li from our center, and the corresponding author is Professor Guihua Shan from our center.

This work was supported by the National Key Research and Development Program of China (Grant No.2023YFB3002500) and the National Natural Science Foundation of China (No.62202446).

The workflow of ParamsDrag model

Interactive analysis of cosmic simulation parameters

Related work

Guan Li, Yang Liu, Guihua Shan*, Weiqun Cao, Ko-Chih Wang. ParamsDrag: Interactive Parameter Space Exploration via Image-Space Dragging, IEEE Visualization,13-17, Oct. 2024