Public Abstract

DE-SC0025347: DyGenAI: Dynamic Generative Artificial Intelligence for Prediction and Control of High-Dimensional Nonlinear Complex Systems

Award Status: Active

Institution: The Trustees of Columbia University in the City of New York (Morningside Campus), New York, NY
UEI: F4N1QNPB95M4
DUNS: 049179401

Most Recent Award Date: 09/05/2024
Number of Support Periods: 1
PM: Spotz, William

Current Budget Period: 09/01/2024 - 08/31/2025
Current Project Period: 09/01/2024 - 08/31/2027
PI: Du, Qiang

Supplement Budget Period: N/A

Public Abstract

Generative artificial intelligence (AI) extends beyond its success in image synthesis, proving itself a powerful uncertainty quantification (UQ) technique through its capability of learning complex high-dimensional probability distributions. Generative AI holds significant potential to enhance decision-making within the U.S. Department of Energy (DOE) when integrated as an outer-loop module with physics-based models or AI foundation models. However, there is a notable research gap given that current generative AI models are mostly static, whereas the DOE’s missions, such as energy distribution and extreme event monitoring, require dynamic generative models capable of efficiently and effectively managing complex, dynamic environments. The static generative AI models, trained on historical data, are inherently limited in processing streaming data and adapting to the real-time changes that characterize many DOE mission-critical areas. This project will develop a dynamic generative artificial intelligence (DyGenAI) paradigm to transform the generative AI methodology from a static paradigm into a computationally efficient dynamic paradigm. The specific project objectives are: (1) to develop dynamic generative AI models capable of integrating physical principles and observational data to predict evolutionary trajectories of high-dimensional and non-linear dynamical systems; (2) to extend our dynamic generative AI models to estimate unobservable states, facilitating their integration as feedback within the control loop to enhance control actions based on partial observations; (3) to scale our dynamic generative AI models on the Oak Ridge Leadership Computing Facility (OLCF)’s supercomputers using exascale parallel-in-time methods to empower real-time prediction and control of complex dynamical systems. The broad impact of this project extends from advancing fundamental AI methodology to addressing critical areas of DOE’s mission. Specific examples include: (1) Enhancing trustworthiness of AI foundation models. DyGenAI can serve as an outer-loop module surrounding AI foundation models, leveraging observational data to correct potential model errors and thereby enhancing the reliability of AI model predictions. (2) Improving energy system’s reliability. DyGenAI can be used to continuously learn and adapt from real-time data streams, supporting informed decision-making for energy security and clean energy production. This capability not only enhances the trustworthiness, safety, and security of AI but also improves the reliability of energy systems, informs risk mitigation strategies, and supports climate-resilient decision-making.

Generative artificial intelligence (AI) extends beyond its success in image synthesis, proving itself a powerful uncertainty quantification (UQ) technique through its capability of learning complex high-dimensional probability distributions. Generative AI holds significant potential to enhance decision-making within the U.S. Department of Energy (DOE) when integrated as an outer-loop module with physics-based models or AI foundation models. However, there is a notable research gap given that current generative AI models are mostly static, whereas the DOE’s missions, such as energy distribution and extreme event monitoring, require dynamic generative models capable of efficiently and effectively managing complex, dynamic environments. The static generative AI models, trained on historical data, are inherently limited in processing streaming data and adapting to the real-time changes that characterize many DOE mission-critical areas.

This project will develop a dynamic generative artificial intelligence (DyGenAI) paradigm to transform the generative AI methodology from a static paradigm into a computationally efficient dynamic paradigm. The specific project objectives are: (1) to develop dynamic generative AI models capable of integrating physical principles and observational data to predict evolutionary trajectories of high-dimensional and non-linear dynamical systems; (2) to extend our dynamic generative AI models to estimate unobservable states, facilitating their integration as feedback within the control loop to enhance control actions based on partial observations; (3) to scale our dynamic generative AI models on the Oak Ridge Leadership Computing Facility (OLCF)’s supercomputers using exascale parallel-in-time methods to empower real-time prediction and control of complex dynamical systems.

The broad impact of this project extends from advancing fundamental AI methodology to addressing critical areas of DOE’s mission. Specific examples include: (1) Enhancing trustworthiness of AI foundation models. DyGenAI can serve as an outer-loop module surrounding AI foundation models, leveraging observational data to correct potential model errors and thereby enhancing the reliability of AI model predictions. (2) Improving energy system’s reliability. DyGenAI can be used to continuously learn and adapt from real-time data streams, supporting informed decision-making for energy security and clean energy production. This capability not only enhances the trustworthiness, safety, and security of AI but also improves the reliability of energy systems, informs risk mitigation strategies, and supports climate-resilient decision-making.