Context-aware Neural-accelerated Visualization Pipeline for Big Volumetric Data

The emergence of multiteraflop machines with thousands of processors for scientific computing combined with advanced sensory-based experimentation has heralded an explosive growth of structured and unstructured data in science and engineering fields. Such scientific endeavors require powerful visualization workflows as a fundamental step to understanding and interpreting big data. An ideal data visualization workflow should provide scientists with a fast and intuitive tool to explore data interactively, identify interesting patterns, observe anomalies, and gain rapid insights into their unique discovery and decision- making processes. In many numerical simulations, e.g., geochemical and geophysical processes, multidimensional data attributes are acquired in 3D volumetric space with multiscale, disparate, and time- dependent dimensions. Despite many advances in visualizing massive volumetric datasets, the lack of an efficient data pipeline inhibits users from rapid visual interactions with the data. It is hypothesized that terabytes of 4D volumetric data can be stored, interpolated, and visualized interactively in real-time (i.e., 60 frame-per-second response) on consumer computers. This will be explored by transforming the data from a discrete form to a continuous form using context-aware implicit neural representations that reduce memory consumption by up to three orders of magnitude while keeping the rendering rate at 60 frame-per-second. The networks will be then converted into GPU program to construct a direct volumetric data rendering pipeline. Our technology, dubbed as neural-accelerated volumetric visualization, NAV2, relaxes the requirements of high-performance computing setup for visualizing big volumetric data, and establishes an intelligent, context-aware data rendering pipeline to explore complex scientific data across all scales interactively in real-time. Uncertainty quantification, generalization and extrapolation to data-poor regions are intrinsic features of our NAV2 tool that are highly in-demand in many scientific fields, within BES-DOE research portfolio, in which data acquisition is costly (e.g., nanoscale and microscale science). In phase I of this SBIR proposal, we will investigate the proof of concept for our novel NAV2 data rendering pipeline, specifically using several large-scale datasets with different characteristics generated for geochemical, biological & geophysical processes (e.g., CO2 subsurface flow, climate system modeling, etc.) and reaction kinetics & compositional processes (e.g., complex materials design) using well-known physics-based computational engines.