Public Abstract

DE-SCL0000071: VisTrust: Probabilistic Visualization of Univariate and Multivariate Scalar Data for Trusted Scientific Analysis and Discovery

Award Status: Active

Institution: Oak Ridge National Laboratory (ORNL), Oak Ridge, TN
UEI: ZLHJJ57QA2H8
DUNS: 099114287

Most Recent Award Date: 09/15/2025
Number of Support Periods: 1
PM: Hawk, Carol

Current Performance Period: 09/15/2025 - 09/14/2026
Current Project Period: 09/15/2025 - 09/14/2030
PI: Athawale, Tushar

Public Abstract

The goal of this project is to research and create foundational techniques for the efficient understanding and communication of uncertainty in two- and three-dimensional (2D/3D) visualizations. This project will address the knowledge gap of how uncertainty is propagated in 2D/3D visualization by creating efficient and scalable probabilistic algorithms that can model, track, and convey uncertainty in large data. Research challenges and gaps: Uncertainty in data remains a major obstacle to reliable scientific analysis and discovery. All data from simulations, experiments, and observations have inherent uncertainty due to model and instrument limitations. Furthermore, as large data are reduced and processed, uncertainty is magnified and nonlinearly mapped in visualization, leading to inaccurate data representation and, consequently, hampering analysis and decision-making. Thus, uncertainty in visualization cannot be ignored. The field of uncertainty quantification (UQ) provides techniques to deal with data uncertainty, commonly represented by 1D error bars or box plots. However, there is a significant gap between the UQ and visualization communities. There are no widespread equivalents to 1D error bars to visualize uncertainty in more complex 2D/3D data. To maximize scientific impact, visualizations must convey uncertainty to enable trusted analysis, a requirement listed as a grand challenge by the U.S. Department of Energy’s (DOE’s) 2022 Advanced Scientific Computing Research (ASCR) visualization report (PRD 1). Although multiple efforts have been invested over the past two decades to quantify and convey uncertainty in 2D/3D scalar visualization, they face three major limitations. First, because they lack a theoretical foundation in uncertainty visualization, most existing techniques resort to empirical approaches (e.g., Monte Carlo sampling) that require significant time to converge. Second, existing empirical techniques do not scale well with multidimensional and multivariate data and increasing data size and are therefore often limited to univariate data. Third, they lack approaches for balancing the trade-offs among the storage and computing overheads for visualizing uncertainty. These gaps in state-of-the-art uncertainty visualization research—missing theoretical approaches, inefficiency and limited scalability, and missing trade-off optimization—are significant barriers to trusted scientific analysis. Objective: By bridging UQ and visualization research, this project aims to innovate theoretical and cost-efficient probabilistic techniques for 2D/3D uncertainty visualization that will address the challenges of state-of-the-art empirical techniques to enable scientists to perform trusted and timely analysis of the data. Methods: To achieve the objective of efficient 2D/3D uncertainty visualization for trusted analysis, we propose three approaches. First, we will investigate closed-form derivations of the propagation of (probabilistic) uncertainty in visualization to address the lack of theory in uncertainty visualization and significantly enhance algorithmic accuracy and efficiency over classical empirical techniques. Second, when closed-form visualization uncertainty is not derivable, we will research ML/AI/data-driven techniques to learn/approximate propagation of uncertainty in visualization for enhanced efficiency and scalability. Lastly, our closed-form and ML/AI/data-driven probabilistic algorithms, combined with parallelization and cost-accuracy trade-offs, will overcome the inefficiency, limited scalability, and impracticality of state-of-the-art empirical techniques to provide reliable uncertainty-aware visualization. Impact: At the end of the 5-year funding period, the proposed research will have created a foundation that scientists can use to efficiently visualize uncertainty in 2D/3D data for trusted analysis (VisTrust). A broad range of applications important to DOE ASCR research, including data reduction, data compression, ensemble analysis, ML/AI for science, and streaming data, will benefit from the proposed research. Specifically, our efficient and scalable uncertainty visualization algorithms will significantly reduce the time to assess uncertainty and build trust in results for these applications. Our research, as presented through scientific publications and an open-source implementation, will be the key to overcoming the lack of a theoretical and cost-efficient framework for uncertainty visualization and to making such a framework accessible to a wider audience.

The goal of this project is to research and create foundational techniques for the efficient understanding and communication of uncertainty in two- and three-dimensional (2D/3D) visualizations. This project will address the knowledge gap of how uncertainty is propagated in 2D/3D visualization by creating efficient and scalable probabilistic algorithms that can model, track, and convey uncertainty in large data.

Research challenges and gaps: Uncertainty in data remains a major obstacle to reliable scientific analysis and discovery. All data from simulations, experiments, and observations have inherent uncertainty due to model and instrument limitations. Furthermore, as large data are reduced and processed, uncertainty is magnified and nonlinearly mapped in visualization, leading to inaccurate data representation and, consequently, hampering analysis and decision-making. Thus, uncertainty in visualization cannot be ignored. The field of uncertainty quantification (UQ) provides techniques to deal with data uncertainty, commonly represented by 1D error bars or box plots. However, there is a significant gap between the UQ and visualization communities. There are no widespread equivalents to 1D error bars to visualize uncertainty in more complex 2D/3D data. To maximize scientific impact, visualizations must convey uncertainty to enable trusted analysis, a requirement listed as a grand challenge by the U.S. Department of Energy’s (DOE’s) 2022 Advanced Scientific Computing Research (ASCR) visualization report (PRD 1).

Although multiple efforts have been invested over the past two decades to quantify and convey uncertainty in 2D/3D scalar visualization, they face three major limitations. First, because they lack a theoretical foundation in uncertainty visualization, most existing techniques resort to empirical approaches (e.g., Monte Carlo sampling) that require significant time to converge. Second, existing empirical techniques do not scale well with multidimensional and multivariate data and increasing data size and are therefore often limited to univariate data. Third, they lack approaches for balancing the trade-offs among the storage and computing overheads for visualizing uncertainty. These gaps in state-of-the-art uncertainty visualization research—missing theoretical approaches, inefficiency and limited scalability, and missing trade-off optimization—are significant barriers to trusted scientific analysis.

Objective: By bridging UQ and visualization research, this project aims to innovate theoretical and cost-efficient probabilistic techniques for 2D/3D uncertainty visualization that will address the challenges of state-of-the-art empirical techniques to enable scientists to perform trusted and timely analysis of the data.

Methods: To achieve the objective of efficient 2D/3D uncertainty visualization for trusted analysis, we propose three approaches. First, we will investigate closed-form derivations of the propagation of (probabilistic) uncertainty in visualization to address the lack of theory in uncertainty visualization and significantly enhance algorithmic accuracy and efficiency over classical empirical techniques. Second, when closed-form visualization uncertainty is not derivable, we will research ML/AI/data-driven techniques to learn/approximate propagation of uncertainty in visualization for enhanced efficiency and scalability. Lastly, our closed-form and ML/AI/data-driven probabilistic algorithms, combined with parallelization and cost-accuracy trade-offs, will overcome the inefficiency, limited scalability, and impracticality of state-of-the-art empirical techniques to provide reliable uncertainty-aware visualization.

Impact: At the end of the 5-year funding period, the proposed research will have created a foundation that scientists can use to efficiently visualize uncertainty in 2D/3D data for trusted analysis (VisTrust). A broad range of applications important to DOE ASCR research, including data reduction, data compression, ensemble analysis, ML/AI for science, and streaming data, will benefit from the proposed research. Specifically, our efficient and scalable uncertainty visualization algorithms will significantly reduce the time to assess uncertainty and build trust in results for these applications. Our research, as presented through scientific publications and an open-source implementation, will be the key to overcoming the lack of a theoretical and cost-efficient framework for uncertainty visualization and to making such a framework accessible to a wider audience.