This work introduces a Scalable Asynchronous Generative Inverse Problem Solver (SAGIPS) for high-performance computing systems. The resulting workflow utilizes an asynchronous ring-allreduce algorithm to transfer the gradients of a Generative Adverserial Network (GAN) across multiple GPUs. Experiments with a scientific proxy application demonstrate convergence quality comparable to traditional methods and near-linear weak scaling. This novel approach holds the potential to significantly advance methods for solving complex, large-scale inverse problems.

This work studies how stresses and momentum flow operate in bound quantum systems, with special attention paid to how stresses induced by the wave function guide particle motion. This study additionally shows how mechanical form factors—which are extracted from the quantum correlation functions measured at JLab and the EIC—can be used to image forces in bound quantum systems. The hydrogen atom is used as a case study, and it is shown that the mechanical form factors allow the Coulomb force law to be mapped out.

The QuantOm has collaboration developed an ultra-fast software package called tiktaalik to perform renormalization group evolution of generalized parton distributions. The code uses finite element methods to make the evolution auto-differentiable, allowing it to be used in AI/ML frameworks, such as those being developed by QuantOm.

Accessing the hidden universe of quarks and gluons is a challenge requiring expertise across multiple disciplines. Theorists provide rigorous mathematical expressions that connect the footprints of this hidden universe to observational data. Experimentalists enable realistic simulations of complex particle reactions at world-class facilities such as Jefferson Lab and the future Electron-Ion Collider. To reconstruct the details of this hidden universe, data scientists orchestrate end-to-end, simulation-based inference. Processing the full simulation pipeline and the large-scale data from these experiments demands high-performance computing expertise. Finally, data-visualization specialists render the reconstructed information as tomographic images of the sub-femtometer-scale world.

Our theory team, consisting of researchers from Jefferson Lab and Argonne National Laboratory, is pushing forward the frontiers of quantum chromodynamics by developing and extending factorization frameworks. These allow quantum correlation functions to be measured in an increasingly large variety of experimental processes at Jefferson Lab and the future Electron Ion Collider, and also allow radiative effects from electromagnetic interactions to be incorporated into theoretical models. Moreover, we are developing and deploying numerical implementations of these frameworks using modern finite element methods, allowing the frameworks to be employed in an auto-differentiable, machine learning-friendly pipeline.

TODO

Our experimental working group, including researchers from Jefferson Lab, is developing a novel, simulation-based inference method in a conceptual Bayesian framework designed for Generative Adversarial Networks (GANs). We develop new techniques to solve ill-posed inverse problems powered by advanced scientific computing and machine learning methods. The conceptual Bayesian GAN framework allows for accessible quantification of statistical uncertainties. The accurate estimation of statistical uncertainties is necessary to reveal systematic effects and bias. These methods provide high-fidelity and expressive extraction of Quantum Correlation Functions that can be seamlessly incorporated into our event-level framework for effective representation of physics experiments.

Our applied mathematics team is developing a next-generation, simulation-based inference methodology that allows accurate event-level data analysis. We develop scalable inverse-problem solvers powered by advanced scientific machine learning coupled with rigorous statistical methodology, advanced sampling, and fast surrogate models that accelerate inference by orders of magnitude. Together, these advances enable the rapid and high-fidelity extraction of QCFs, providing theory with sharper constraints and experiment and detector models with more realistic representations.

Our supercomputing and visualization team is developing a scalable distributed learning framework, along with associated libraries and tools, to overcome the computational challenges of performing quark-gluon correlation function (QCF) inference, which, in turn, enables the femtoscale imaging of nuclei on exascale systems. As such, the team interfaces with applied mathematicians and AI researchers to map, optimize, or refactor the algorithms within the aforementioned framework, libraries, and tools for high performance on exascale systems, including a comprehensive visualization framework or library to support real-time, human-in-the-loop decision-making in machine-learning workflows, focusing on QCFs and their uncertainties.