Innovative method brings faster and more efficient visualization of CDF simulation data

Researchers from the IT4Innovations National Supercomputing Centre at VSB-TUO have created an innovative methodology for visualizing large-scale data from computational fluid dynamics (CFD) simulations using volume rendering and advanced visualization techniques. This method, developer using Karolina supercomputer, allows for rapid processing of CFD simulation outcomes and, when paired with interactive photorealistic visualization, opens up new avenues for analyzing flow simulations. The research received funding from the REFRESH project.

CFD simulations are highly intensive in computational demands and often generate huge terabytes to petabytes of resulting data. Efficient processing of this data is crucial for fast and correct analysis of the results. Particularly with unstructured meshes, post-processing requires significant computational power and memory resources. Currently available tools cannot efficiently process large, time-dependent unstructured data in a reasonable time.

Researchers from IT4Innovations have developed a new workflow to process data that allows the re-sampling of hundreds of time steps on an unstructured mesh with one billion cells (tens of terabytes of data) to a sparse regular grid with a density of 11 billion voxels (3D blocks) in minutes. This process uses thousands of processor cores and advanced algorithms for efficient data preparation, which is a prerequisite for the subsequent interactive visualization.

The whole workflow consists of five key steps. First, data is loaded into memory in parallel to ensure fast upload of large data sets. The data is then evenly distributed across the available computational resources, optimizing hardware utilization, and speeding up processing. Unstructured data is then transformed into a regular grid (voxelisation), which simplifies subsequent processing. After resampling, the data is stored in OpenVDB format, the industry standard for computer graphics. Finally, the data is prepared for high-fidelity visualization using Blender to create visually attractive and detailed output for analysis and presentation.

A unique feature of the developed solution is the GPU-accelerated interactive display of time-dependent simulations through volume rendering using a path-tracing renderer. The volumetric representation of simulated quantities allows for efficient interactive exploration of the phenomena in both space and time interactively.

This new methodology for visualizing large-scale data has been validated on specific simulations, such as commercial aircraft. It took 37 minutes to prepare a one-billion cell CFD simulation with 512-time steps (34 TB of data) for an interactive volumetric rendering with 11 billion voxels. This procedure outperforms current visualization techniques and brings a new level of voxelisation scalability.

The method, tested on the Karolina supercomputer in Ostrava, is an efficient tool for post-processing and visualizing large-scale CFD data, allowing researchers and engineers to analyze and visualize simulations faster, even for models with billions of cells. In the future, the research team will focus on implementing a hierarchical approach to reduce the memory demand while maintaining the visual quality of the output, and on further data reduction using neural networks.