Rochester-Area Collaboration Drives Global Scientific Progress: LLE and RIT Look to Unlock the Secrets of Extreme Matter Using AI/ML

A photograph of a group of people standing on a staircase.

RIT Professor Qi Yu with HEDP Theory Group members, Suxing Hu, Deyan Mihaylov, Valentin Karasiev, and Rati Goshadze standing in The Shed's atrium at the Rochester Institute of Technology on Monday, November 25, 2024.

The 2024 Nobel Prize in Physics recognized the inventors of neural networks, highlighting the revolutionary impact of artificial intelligence (AI) and machine learning (ML) across scientific disciplines. AI is proving to be a game changer in the realm of high-energy-density (HED) physics—a field exploring matter under extreme conditions. A dynamic collaboration between the University of Rochester’s Laboratory for Laser Energetics (LLE) and Rochester Institute of Technology (RIT) is at the forefront of this exciting intersection of AI and HED physics.

Tackling Cosmic Complexities

HED physics investigates matter under pressures millions of times greater than Earth’s atmosphere and at temperatures rivaling those found at the cores of stars. These extreme conditions, found in astrophysical objects and in fusion experiments, present unique challenges for researchers. “To unravel the underlying physics and trend, we need AI/ML experts like Professor Qi Yu at RIT’s Machine Learning and Data Intensive Computing (Mining) Lab in the School of Information to help ‘mine’ the ‘precious’ HED data from complicated and expansive first-principles calculations and experiments,” explains Distinguished Scientist Suxing Hu, who leads LLE’s High-Energy-Density Physics (HEDP) Theory Group. The complexity of modeling such extreme environments, the computational expense of large-scale simulations, and the need to extract meaningful insights from vast, noisy datasets have long been hurdles in the field. AI and ML techniques offer powerful solutions to these challenges, enabling more-efficient data analysis, optimized simulations, and improved predictions of complex behaviors.

A Fusion of Expertise

The collaboration between LLE and RIT exemplifies how local expertise can drive global scientific progress, built on a foundation of shared community ties and evolving professional goals. Hu and Yu first became acquainted nearly a decade ago and over time, their interactions grew into a deeper appreciation for each other’s work. This foundation paved the way for a formal partnership four years ago, when an opportunity arose to unite their respective strengths in HED physics and AI.

Hu notes, “This collaboration between LLE and RIT was built naturally out of the needs and by the dedicated efforts of experts based in the Rochester region.” Similarly, Yu recognizes the synergy that has resulted from this ideal and dynamic partnership: “The collaboration with Dr. Hu’s HEDP Theory Group at LLE offers an invaluable opportunity to demonstrate how AI/ML can address long-standing challenges in HED physics. This initiative aligns with the core research theme of the Mining Lab at RIT, which emphasizes the development of innovative, resource-efficient AI systems to advance scientific discovery.”

Equation of State: Mapping the Extreme

One of the team’s notable achievements involves using AI/ML to construct global equation-of-state (EOS) models. EOS describes how a material’s internal energy and pressure change with density and temperature—critical information for understanding phenomena from planetary structures to fusion energy experiments. EOS is fundamental to understanding the behavior of matter, from determining the size of stars to designing fusion targets for clean energy production.

Traditional EOS measurements and calculations are limited to specific conditions, making comprehensive modeling challenging. To combat these challenges, the LLE–RIT team developed a novel deep regression model for energy and pressure (DREP) that can accurately predict EOS values across a wide range of densities and temperatures, even with limited input data (Fig. 1). This model is now being used in LLE’s radiation-hydrodynamics codes to predict material behaviors across a wide range of conditions with unprecedented accuracy—in other words, to simulate inertial confinement fusion experiments.

Graph showing pressure versus temperature with a number of colorful dashed lines and circular data points. — Figure 1. DREP prediction results: density values included in the training data (solid curves), unseen density values (dashed curves), and ground-truth points (circles).

Pushing the Boundaries of Density Functional Theory

The collaboration is now setting its sights on improving the accuracy of density-functional-theory (DFT) calculations for HED systems. DFT calculates the interactions between particles using quantum mechanical modeling to predict the material’s properties. By developing a continual active learning and testing (CALT) framework (Fig. 2), the team aims to discover a universal exchange-correlation (XC) functionals that can dramatically enhance DFT simulations under extreme conditions.

CALT aims to address several core challenges by:

Learning a highly accurate XC potential function from sparse training data from expensive wave-function–method computations
Overcoming constant domain shifts to ensure that the learned function is universal and can generalize to diverse elements
Incorporating important physical laws to discover physics-consistent functions

This work could lead to a paradigm shift in HED physics research, enabling more-efficient and accurate computations of material properties under extreme conditions. The implications range from better understanding astrophysical phenomena to accelerating the development of fusion energy technologies.

Illustration showing the framework in two vertical panels. — Figure 2. Overview of the CALT framework that highlights three main thrusts: (1) Thrust I focuses on both accurate calculations and basic DFT computations using temperature-dependent XC functionals for various atomic elements under extreme HED conditions. The former method will provide sparse but accurate ground-truth data, while the latter can give plenty of mid-fidelity data for pre-training a foundation model of XC functionals. (2) Thrust II explores data-efficient fine-tuning of the foundation model with guaranteed accuracy. (3) Thrust III performs continual learning to accumulate the global knowledge over a sequence of atoms toward learning a universal XC functional.

A Model for Interdisciplinary Innovation

The LLE–RIT collaboration demonstrates the power of partnership in tackling complex scientific challenges. As Professor Yu notes, ”RIT has a longstanding tradition of fostering interdisciplinary research, and the collaboration with LLE, which integrates diverse expertise from distinct scientific disciplines, exemplifies this commitment.”

By bridging the gap between AI and HED physics, this partnership not only advances our understanding of matter under extreme conditions but also paves the way for future breakthroughs in fields ranging from astrophysics to fusion energy. As AI continues to evolve, its application in HED physics promises to unlock new frontiers in our quest to understand the fundamental nature of the universe.

HEDP Theory Group and RIT Group standing in front of Innovation Hall at the Rochester Institute of Technology. — HEDP Theory Group members, Rati Goshadze, Deyan Mihaylov, Suxing Hu, and Valentin Karasiev with RIT Professor Qi Yu and RIT students: Dayou Yu and Deep Shankar Pandey in front of the Innovation Hall at the Rochester Institute of Technology on Monday, November 25, 2024.

Focus Points

A collaborative LLE–RIT team is using artificial intelligence (AI) and machine learning (ML) to address several challenges in high-energy-density (HED) physics.
The team has achieved its first goal: building highly accurate and thermodynamically consistent equation-of-state models.
Further collaboration is underway toward learning the “holy grail” of exchange-correlation functionals for improving density-functional-theory simulations of HED matter.

A version of this article appears in Issue 7 of LLE In Focus, the magazine of the University of Rochester’s Laboratory for Laser Energetics.