Compression and Ray-March Rendering Using Implicit Neural Representations of Data Values and Domain Extent

Unlike traditional approaches, neural representations can compress unstructured data from aerospace simulations by factors of 200x while retaining sufficient numerical quality for visualisation. These work by approximating the underlying field, mapping coordinates to values via compact neural networks. A key practical limitation, however, is that this mapping does not encode the spatial extent of the original dataset domain, which almost always leads to erroneously extrapolated artefacts in rendered images produced by implicit ray-march rendering. Previous compression studies avoid this by separately storing the original coordinates to reconstruct an explicit version of the data for post-compression visualisation. This approach, however, reduces the effective compression of both coordinates and values to a negligible amount. As a solution, we propose a dual-network framework that jointly encodes both data values and domain extent within a fully implicit, compressive neural representation. Our approach uses a compressed neural signed distance function to constrain data to its true bounds, enabling implicit rendering without intermediate decompression. We validate this approach using our hybrid sphere-tracing and ray-marching method for isosurface rendering, and demonstrate its effectiveness — along with several optimisations — using a CUDA proof-of-concept viewer.