Prof. Dr. Rüdiger Westermann

We have addressed the question whether neural volume representation networks can effectively compress and efficiently reconstruct volumetric data sets, and whether they can even be used for temporal reconstruction tasks. We have proposed a novel design of such networks using GPU tensor cores to integrate the reconstruction seamlessly into on-chip raytracing kernels. Comparisons to alternative implementations and compression schemes show significantly improved performance at compressions rates that are on par with those of the most effective competitors. For time-varying fields, we have proposed a solution that builds upon latent-space interpolation to enable random access reconstruction at arbitrary granularity and high reconstruction fidelity.

We have built upon our work to radically change the way so-called de-homogenization is performed, to efficiently design high-resolution load-bearing structures. Our new approach builds upon streamline-based parametrization of the design domain, but now uses direction fields that are obtained from homogenization-based topology optimization, i.e., where an optimal distribution of continuous material values is performed instead of generating a binary design. Streamlines in these fields are then converted into a quad-dominant mesh, and the widths of the grid edges are adjusted according to the density and anisotropy of the optimized cells. In a number of numerical examples, we have demonstrated the mechanical performance and regular appearance of the resulting structural designs.

Within the Collaborative Research Center "Waves to Weather", funded by the German Research Foundation, we work together with meteorologist to develop efficient visual analysis techniques for multi-parameter data, i.e., sets of data points with potentially many associated parameter values per data point. We have proposed a scalable GPU realization of parallel coordinates building upon 2D pairwise attribute bins, to significantly reduce the number of lines to be rendered and, thus, make parallel coordinates applicable to weather forecast ensembles comprising billions of data points, each carrying multiple prognostic floating-point variables like temperature, precipitation and pressure, plus spatial and simulation input variables.

Can a volumetric field or a color transfer function used to visualize this field be reconstructed from few images of the field. To achieve this, we have presented a differentiable volume rendering solution that provides differentiability of all continuous parameters of the volume rendering process. This differentiable renderer is used to steer the parameters towards a setting with an optimal solution of a problem-specific objective function. We have tailored the approach to volume rendering by enforcing a constant memory footprint via analytic inversion of the blending functions. This makes it independent of the number of sampling steps through the volume and facilitates the consideration of small-scale changes.

We have introduced a new topology optimization approach for structural design that is guided by stresslines in a solid under load. This is particularly appealing from a computational perspective, since it avoids iterative finite element optimizations. By regularizing the thickness of each stressline using derived strain energy measures, stiff structural layouts are generated in a highly efficient way. We have then proposed a method that uses the resulting structures as initial density fields in density-based topology optimization, and demonstrate that by using a stressline guided density initialization in lieu of a uniform density field, convergence issues in density-based topology optimization can be significantly relaxed at comparable stiffness of the resulting structural layouts.

We have used advanced GPU shader functionality to generate a focus+context mesh renderer that highlights the elements in a selected region and simultaneously conveys the global mesh structure and deformation field. Therefore, we have developed a new method to construct a sheet-based level-of-detail hierarchy and smoothly blend it with volumetric information. A gradual transition from edge-based focus rendering to volumetric context rendering is performed, by combining fragment shader-based edge and face rendering with per-pixel fragment lists.

With 3D-TSV, we have introduced a visual analysis tool for the exploration of the principal stress directions in 3D solids under load. 3D-TSV provides a modular and generic implementation of key algorithms required for a trajectory-based visual analysis of principal stress directions, including the automatic seeding of space-filling stress lines, their extraction using numerical schemes, their mapping to an effective renderable representation, and rendering options to convey structures with special mechanical properties.

Within the Collaborative Research Center "Waves to Weather", funded by the German Research Foundation, we continue our research at the interface between meteorology and computer science. In the second funding period, one key focus is on Deep Weather, i.e., the use of artificial neural networks for data inference related to weather forecasting. First results showing the potential of neural networks for down-scaling, i.e., inferring high-resolution wind fields from given low-resolution wind fields and their relation to high-resolution topography and boundary conditions, were presented at the ECMWF Copernicus meeting 2019.  

Over the last two decades, we have shaped the landscape of volume rendering, by our developments of GPU-based techniques, effective acceleration structures, and efficient compression schemes. Now its time for another innovation in this field: The use of Deep Learning to reduce the number of samples that are needed to render a volumetric data set.

Our results demonstrate that artificial neural networks can learn to infer the geometric properties of isosurfaces in volumetric scalar fields, rather than learning a specific isosurface (images below show network-based inference - right - from low resolution images at 1/16 of the samples -left. This enables to significantly reduce the number of volumetric samples that need to be taken to generate an image of an isosurface. Our research now aims at extending this work towards direct volume rendering, and using neural networks to guide renderers towards specific data features.

In an elaborate evaluation, we have investigated CPU and GPU rendering techniques for large transparent 3D line sets. We compare accurate and approximate techniques using a number of benchmark data sets. Besides optimized implementations of rasterization-based object- and image-order methods, we consider Intel’s OSPRay CPU ray-tracing framework, a GPU ray-tracer using NVIDIA’s RTX ray-tracing interface through the Vulkan API, and voxel-based GPU line ray-tracing.