Public Abstract

DE-SC0023195: A Scientist-in-the-Loop Data Analytics Framework for Intelligent Simulation Model Tuning and Validation

Award Status: Active

Institution: New Jersey Institute of Technology, Newark, NJ
UEI: SGBMHQ7VXNH5
DUNS: 075162990

Most Recent Award Date: 08/30/2022
Number of Support Periods: 1
PM: Nasiri, Shaima

Current Budget Period: 09/01/2022 - 05/31/2024
Current Project Period: 09/01/2022 - 05/31/2024
PI: Dasgupta, Aritra

Supplement Budget Period: N/A

Public Abstract

A Scientist-in-the-Loop Data Analytics Framework for Intelligent Simulation Model Tuning and Validation PI: Aritra Dasgupta, Assistant Professor Co-PI: Chase Wu, Professor Department of Data Science, New Jersey Institute of Technology (NJIT) Scientific advances in the big data era are contingent upon the rate at which insights are produced by processing and analyzing data generated by model-based simulations. The development of such models must iterate through complex parameterization and tuning. For scientists, it is a labor-intensive, time-consuming process requiring careful consideration and calibration of various trade-offs using their domain knowledge. Selecting the best model and identifying critical parameters among the combinatorically large number of possibilities is just as challenging as looking for a needle in a haystack. To address this challenge, we propose to develop human-in-the-loop visual analytic techniques that enable scientists to proactively supervise, modify, and validate model-based simulations at scale using intelligent data-driven methods. These techniques combine the best of both worlds of intelligent automation, by leveraging the predictive power of machine learning, and human expertise, by developing expressive visualization techniques that intuitively communicate the causes behind good and bad model outcomes. The project's vision is to transform both the process and the outcome of the state-of-the-art scientific model-based simulation processes by providing: i) intelligent steering services that guide scientists in optimally tuning models. ii) advanced visualization techniques allowing scientists to focus on reasoning about factors affecting model performance, and iii) interactive user interfaces that drastically reduce scientists' investment of time in hypothesis generation, experiment configuration, and model validation. We will instantiate the proposed solution in collaboration with climate scientists at national labs by initially focusing on faithful simulation of clouds, which is critical for scientists' understanding and prediction of the global climate and climate change-related patterns.

A Scientist-in-the-Loop Data Analytics Framework for Intelligent Simulation Model Tuning and Validation
PI: Aritra Dasgupta, Assistant Professor
Co-PI: Chase Wu, Professor
Department of Data Science, New Jersey Institute of Technology (NJIT)

Scientific advances in the big data era are contingent upon the rate at which insights are produced by processing and analyzing data generated by model-based simulations. The development of such models must iterate through complex parameterization and tuning. For scientists, it is a labor-intensive, time-consuming process requiring careful consideration and calibration of various trade-offs using their domain knowledge. Selecting the best model and identifying critical parameters among the combinatorically large number of possibilities is just as challenging as looking for a needle in a haystack.

To address this challenge, we propose to develop human-in-the-loop visual analytic techniques that enable scientists to proactively supervise, modify, and validate model-based simulations at scale using intelligent data-driven methods. These techniques combine the best of both worlds of intelligent automation, by leveraging the predictive power of machine learning, and human expertise, by developing expressive visualization techniques that intuitively communicate the causes behind good and bad model outcomes.

The project's vision is to transform both the process and the outcome of the state-of-the-art scientific model-based simulation processes by providing: i) intelligent steering services that guide scientists in optimally tuning models. ii) advanced visualization techniques allowing scientists to focus on reasoning about factors affecting model performance, and iii) interactive user interfaces that drastically reduce
scientists' investment of time in hypothesis generation, experiment configuration, and model validation.

We will instantiate the proposed solution in collaboration with climate scientists at national labs by initially focusing on faithful simulation of clouds, which is critical for scientists' understanding and prediction of the global climate and climate change-related patterns.