Deep Hierarchical Super Resolution for Scientific Data

Abstract

Scene representation networks (SRNs) have been recently proposed for compression and visualization of scientific data. However, state-of-the-art SRNs do not adapt the allocation of available network parameters to the complex features found in scientific data, leading to a loss in reconstruction quality. We address this shortcoming with an adaptively placed multi-grid SRN (APMGSRN) and propose a domain decomposition training and inference technique for accelerated parallel training on multi-GPU systems. We also release an open-source neural volume rendering application that allows plug-and-play rendering with any PyTorch-based SRN. Our proposed APMGSRN architecture uses multiple spatially adaptive feature grids that learn where to be placed within the domain to dynamically allocate more neural network resources where error is high in the volume, improving state-of-the-art reconstruction accuracy of SRNs for scientific data without requiring expensive octree refining, pruning, and traversal like previous adaptive models. In our domain decomposition approach for representing large-scale data, we train an set of APMGSRNs in parallel on separate bricks of the volume to reduce training time while avoiding overhead necessary for an out-of-core solution for volumes too large to fit in GPU memory. After training, the lightweight SRNs are used for realtime neural volume rendering in our open-source renderer, where arbitrary view angles and transfer functions can be explored. A copy of this paper, all code, all models used in our experiments, and all supplemental materials and videos are available at https://github.com/skywolf829/APMGSRN.

A Prediction‐Traversal Approach for Compressing Scientific Data on Unstructured Meshes with Bounded Error

“…Dimension‐reduction‐based compressors (e.g., TTHRESH [BRLP19]) reduce data dimensions by techniques such as higher‐order singular vector decomposition (HOSVD). Recently, neural networks have been widely used to reconstruct scientific data, such as autoencoders [LDZ*21, ZGS*22], super‐resolution networks [WGS*23, HZCW22], and implicit neural representations [XTS*22, LJLB21, WHW22, MLL*21, SMB*20]. Yet, most neural compressors do not offer explicit pointwise error control for scientific applications.…”

Super-resolution analysis via machine learning: a survey for fluid flows

“…For super resolution in which the dimension of the output R n is larger than that of the input R m , there are several ways to treat the difference of the dimensions between the input and the output. For example, the upsampling can be used inside a network to expand the dimension [79]. One can also implement a resize or interpolation function for the input data to align the size with that of the output [29,54,80].…”

Trajectory super-resolution methods

“…These methods learn the complex correspondence between low and high-resolution data, and have shown remarkable performance. [2,[13][14][15][16][39][40][41] Although widely used, several limitations still remain for deep learning-based super-resolution methods. First, current methods learn a deterministic one-to-one mapping between high and low-resolution data pairs.…”