The field of hadronic physics is poised for significant advancements, driven by upcoming experimental programs at Jefferson Lab's 12 GeV facility and the future Electron-Ion Collider at Brookhaven National Laboratory. These experiments will produce vast amounts of data, necessitating sophisticated analytical methods to bridge experimental observations and the underlying theoretical frameworks of quantum chromodynamics (QCD). This paper introduces a generative AI-based surrogate modeling framework to address the computational and analytical challenges of traditional theory-driven approaches in event-level analysis. Specifically, we evaluate and compare the effectiveness of generative AI models, including Generative Adversarial Networks (GANs) and Diffusion Models (DMs), as surrogates for event generation. By replacing complex quantum correlation function (QCF) models with generative AI-based surrogates, our proposed framework enables efficient generation of synthetic events directly from QCF parameters. Moreover, this approach facilitates simulation-based parameter inference by learning the mappings between QCF parameters and observables, bypassing the gradient calculation challenges when directly incorporating the QCF model into solving the QCD inverse problems. Through a proof-of-concept QCF study, we demonstrate the capability of generative AI-based surrogates as an important tool to accurately replicate theoretical QCF models, which can be easily incorporated into simulation-based inference to reconstruct QCF parameters from observables.

The “acceptance problem” in high energy physics (HEP) refers to the challenge of accurately modeling detector acceptance to ensure the precision of measurements. This study explores the application of Generative AI to address the acceptance problem in HEP. By training a Generative Adversarial Network (GAN) on simulated detector data (pseudo-data), we demonstrate its capability to learn detector responses and generate synthetic data that closely match measured distributions. A key component of our methodology is a custom generator loss function that incorporates physics-informed principles to improve training. This custom loss function penalizes deviations from the true distribution of event components, ensuring that the generated samples adhere to the underlying physics. Additionally, we trained a binary classifier to distinguish between different topological states (measured and unmeasured events) within the generated Monte Carlo pseudodata, further refining the model's accuracy. Our approach preserves correlations between kinematic variables across multiple dimensions, providing an accurate representation of the underlying physics. Validation with Monte Carlo pseudodata demonstrates the method's ability to recover true distributions even in regions with limited detector sensitivity, establishing a solid foundation for applying our framework to real experimental data. Our results highlight the feasibility and advantages of using generative AI in HEP, paving the way for broader applications in the field.

Correctly mapping the experimental data to quantum probability distributions is a critical step to characterize nucleon structure and the emergence of hadrons, in terms of quark and gluon degrees of freedom. Since the actual parameters of interest are not directly measurable, but instead are inferred from experimental observables, this is fundamentally an inverse problem of recovering parameters from observables. In addition to the well-known challenges such as ill-posedness in general inverse problems, an application specific issue here is that the experimental data are observed on kinematics bins which are usually irregular and varying. In this paper, to address this ill- defined, varying observable space problem, we represent the observables together with their kinematics bins as an unstructured, high-dimensional point cloud. We incorporate a permutation invariant neural network framework to handle the observables in unstructured and unordered point cloud representations. We incorporate the point cloud representation into the Variational Autoencoder Inverse Mapper (VAIM) framework. The point cloud-based VAIM (PC-VAIM) enables the underlying deep neural networks to learn how the observables are distributed across kinematics. We demonstrate the effectiveness of PC-VAIM on a toy inverse problem, and then on constructing the inverse function mapping Quantum Correlation Functions (QCF) to observables in a Quantum Chromodynamics (QCD) analysis of nucleon structure.

We apply generative adversarial network (GAN) technology to build an event generator that simulates particle production in electron-proton scattering that is free of theoretical assumptions about underlying particle dynamics. The difficulty of efficiently training a GAN event simulator lies in learning the complicated patterns of the distributions of the particles physical properties. We develop a GAN that selects a set of transformed features from particle momenta that can be generated easily by the generator, and uses these to produce a set of augmented features that improve the sensitivity of the discriminator. The new Feature-Augmented and Transformed GAN (FAT-GAN) is able to faithfully reproduce the distribution of final state electron momenta in inclusive electron scattering, without the need for input derived from domain-based theoretical assumptions. The developed technology can play a significant role in boosting the science of existing and future accelerator facilities, such as the Electron-Ion Collider.

Event generators in high-energy nuclear and particle physics play an important role in facilitating studies of particle reactions. We survey the state of the art of machine learning (ML) efforts at building physics event generators. We review ML generative models used in ML-based event generators and their specific challenges, and discuss various approaches of incorporating physics into the ML model designs to overcome these challenges. Finally, we explore some open questions related to super-resolution, fidelity, and extrapolation for physics event generation based on ML technology.

Inverse problems - using measured observations to determine unknown parameters - are well motivated but challenging in many science and engineering problems. In this paper, we propose an end-to-end deep learning framework, the Variational Autoencoder Inverse Mapper (VAIM), as an autoencoder-based neural network architecture for inverse problems. The encoder and decoder neural networks approximate the forward and backward mapping, respectively, and a variational latent layer is incorporated into VAIM to learn the posterior parameter distributions with respect to given observables. We demonstrate the effectiveness of VAIM for several toy inverse problems, with both finite and infinite solutions, and for constructing the inverse function mapping quantum correlation functions to observables in a Quantum Chromodynamics analysis of nucleon structure.

The recently proposed generative adversarial network (GAN) based event generator, the Feature Augmented and Transformed GAN (FAT-GAN), has shown impressive capability of reproducing inclusive electron-proton scattering events at a given collision energy. In contrast, many practical applications require the event generator to have the flexibility of allowing users to specify the reaction energy as an input to produce the corresponding synthetic events. In this work, we extend the FAT-GAN framework by conditioning the component neural networks according to the given reaction energy. We demonstrate that this model, referred to as cFAT-GAN, can reliably produce inclusive event feature distributions and correlations for a continuous range of reaction energies by automatically interpolating and extrapolating from a set of trained energies. We employ a continuous energy feature representation to enable the networks to organically learn the distribution relationships between different reaction energies, laying the groundwork for accessing events at untrained energies.This continuous conditional energy provides a degree of versatility to the cFAT-GAN for its further development as a significant research tool in high-energy and nuclear physics.

In this paper, we evaluate methods of aggregating personal sensor data to estimate the degree of group stress. We compare the estimation accuracy of two aggregation methods, one without and one with personal distribution scaling. As a result of applying the data of 9 groups of people (105 people) to the methods, it is shown that the no-scaling method successfully estimates the degree of group stress better than the other (average error = 1.19, r = 0.964). At the conclusion of discussion, it is suggested that the method without scaling is valid because both the group stress state and method reflect the interaction among people.

A highly practical use case of pedestrian-track analysis by using LiDAR is presented in this paper. Many problems are caused by heavy crowdedness in the management of public facilities, i.e., shopping malls, airports, and so on. One solution is crowdedness equalization by controlling pedestrian flow. We conducted two experimental demonstrations at technical exhibitions to find that factors that determine pedestrian flow. Pedestrian tracks were obtained at an exhibition in 2013 by using a LiDAR-based pedestrian-tracking system first. As a result, new knowledge was gained; the layout of a technical exhibition should be designed to bend the path of a pedestrian flow toward areas where their attention is desired to be. The layout of an exhibition in 2014 was designed to bend the pedestrian path many times so that pedestrians' attention was located diversely. Therefore, pedestrian tracks were successfully obtained; as a result, it was confirmed that crowdedness was successfully equalized.

A temporal network consists of a time series of interaction events, each of which is defined by a triplet composed of the indices of two nodes and the time of the event. Mapping a temporal network to a more tractable static network is often useful. A mapping method was recently proposed on the basis of the so-called transfer entropy (G. V. Steeg and A. Galstyan, in Proc. the 21st Int. Conf. WWW, p.509, 2012). In the proposed method, one generates the directed network of influence in which a directed link represents the causal relationship between activity patterns at two nodes. However, the significance of the inferred links and the sensitivity of results to the parameter values are still unclear. We propose a bootstrap sampling method to statistically configure the directed network of influence. We apply our method to the face-to-face interaction logs between office workers in Japanese companies.

A badge-shaped sensor and feedback system were developed. This system makes it possible to study human and organizational behavior in an office. By prompting changes in workers' behavior, it improves productivity as well as helps to solve individual's problems. Aiming to improve the productivity and quality of project management, the system was applied at a software development company. The results show that the system can improve the human communication process, as well as motivate workers by giving them the chance to reflect on their work styles and to help their colleagues.

Visualization of the actual conditions of an organization is a very challenging issue. We propose a new system called a Business Microscope that senses the activities of people in an organization and provides visual feedback to users. We use name-tag shaped sensor nodes to measure face-to-face interaction between employees. The massive amount of data collected by the sensor-network terminal are signal-processed by the server and displayed as an organization's topographic map that displays the frequencies of organizational activities. To depict the organization's topographic map, our system creates a novel relation tree using the interaction data from all pairs of members. In this kind of map, some groups in the organization hierarchically form islands. Members in those islands who have relationships with many others form mountains that are plotted with contours. We can comprehend the actual conditions of organizations from our topographic map. We tested our technique in several experiments to evaluate this system.

"Business Microscope" is our sensornet application in the age of knowledge, which visualizes knowledge workers' interactions by sensing their face-to-face communications. Due to the limitation of energy consumption of sensor nodes and privacy concerns, very short (0.1s) intermittently sensed (10s interval) noise-like signals called siglet is used to for detection task. To detect the speaker from the limited input, "self" vs "others" classification problem is introduced. For this new classification problem, new classifier called AdaBoost LVQ is studied to explore the application of AdaBoost to reduce the error rate of the conventional classifier with strictly limited inputs. As a result, AdaBoost LVQ achieved highest recognition accuracy of 96.45% with 19.86% error rate improvement relative to best conventional classifier.

"Business Microscope" is a tool which provides knowledge workers with a bird-eye view of their daily communication. To meet the problem of the energy consumption of sensor nodes and privacy concerns for wearers and non-wearers, "siglet" sensing is proposed. Siglet sensing is a way to capture very short and noise-like signals by sensors operating on a low duty ratio. To extract the useful information on workers' communication, speech siglet detection is studied. The LBG trained speech and workplace nonspeech models with Mel frequency cepstrum coefficients (MFCCs) as feature vectors are utilized. A hierarchical pruning technique is studied to reduce the calculation cost of the matching process to nearly 25% and refine the classification accuracy. Our approach achieved average speech and nonspeech classification accuracy of 99.96% on 0. Is long test siglets.

Phonetics and prosody are two important factors in automatic language identification. Prosodic HMMs enable language identification systems to use prosodic information in a similar manner to phonetic HMMs. The paper describes how to create prosodic HMMs and implement them in language identification systems. Linear discriminant analysis of the likelihood and N-gram scores of prosodic segment recognition realizes fast and reliable language identification. Moreover, combining prosodic HMMs with phonetic HMMs improves system performance. In this framework, feature normalization techniques that were originally developed for robust speech recognition can be applied to phonetic and prosodic features. Language identification accuracy increases using these techniques in clean and noisy environments.