Both humans and machine learning models learn from experience, particularly in safety- and reliability-critical domains. While psychology seeks to understand human cognition, the field of Explainable AI (XAI) develops methods to interpret machine learning models. This study bridges these domains by applying computational tools from XAI to analyze human learning. We modeled human behavior during a complex real-world task -- tuning a particle accelerator -- by constructing graphs of operator subtasks. Applying techniques such as community detection and hierarchical clustering to archival operator data, we reveal how operators decompose the problem into simpler components and how these problem-solving structures evolve with expertise. Our findings illuminate how humans develop efficient strategies in the absence of globally optimal solutions, and demonstrate the utility of XAI-based methods for quantitatively studying human cognition.

We're finally reading the secrets of Herculaneum's lost library A whole library's worth of papyri owned by Julius Caesar's father-in-law were turned to charcoal by the eruption of Vesuvius. Deep within a particle accelerator, theoretical physicist Giorgio Angelotti is hard at work. He sets a black cylinder on a mount, bolts it down, then runs through some safety checks before retreating from the chamber, known as "the hatch". "You have to be sure there's no one in the hatch before you close the door," he says. That's because he is about to blast the sample with a super-powerful beam of X-rays.

We present the first language-model-driven agentic artificial intelligence (AI) system to autonomously execute multi-stage physics experiments on a production synchrotron light source. Implemented at the Advanced Light Source particle accelerator, the system translates natural language user prompts into structured execution plans that combine archive data retrieval, control-system channel resolution, automated script generation, controlled machine interaction, and analysis. In a representative machine physics task, we show that preparation time was reduced by two orders of magnitude relative to manual scripting even for a system expert, while operator-standard safety constraints were strictly upheld. Core architectural features, plan-first orchestration, bounded tool access, and dynamic capability selection, enable transparent, auditable execution with fully reproducible artifacts. These results establish a blueprint for the safe integration of agentic AI into accelerator experiments and demanding machine physics studies, as well as routine operations, with direct portability across accelerators worldwide and, more broadly, to other large-scale scientific infrastructures.

A zettabyte is a trillion gigabytes. That's a lot--but, according to one estimate, humanity will produce a hundred and eighty zettabytes of digital data this year. It all adds up: PowerPoints and selfies; video captured by cameras; electronic health records; data retrieved from smart devices or collected by telescopes and particle accelerators; backups, and backups of the backups. Where should it all go, and how much of it should be kept, and for how long? These questions vex the computer scientists who manage the world's storage. For them, the cloud isn't nebulous but a physical system that must be built, paid for, and maintained.

Control of the 6D (x,y,z,p x,p y,p z) phase space distribution of beams is important for all particle accelerators over a wide span of applications including beam - based imaging for material science, accelerator - based light sources, and plasma wakefield - based accelerators. At all large high intensity beam particle accelerators, such as the Los Alamos Neutron Science Center (LAN - SCE) [1] or the Spallation Neutron Source (SNS) [2], the beam's phase distribution must be controlled for proper acceleration and to prevent beam lo ss by matching to the accelerator's magnetic focusing lattice. At particle accelerator - based free electron laser (FEL) light sources, such as the Linac Coherent Light Source [3], the European X - ray FEL [4], and the Swiss FEL [5] the beam's phase space defi nes the properties of the generated light and must be adjusted between different experiments.

Complex dynamical systems, such as particle accelerators, often require complicated and time-consuming tuning procedures for optimal performance. It may also be required that these procedures estimate the optimal system parameters, which govern the dynamics of a spatiotemporal beam -- this can be a high-dimensional optimization problem. To address this, we propose a Classifier-pruned Bayesian Optimization-based Latent space Tuner (CBOL-Tuner), a framework for efficient exploration within a temporally-structured latent space. The CBOL-Tuner integrates a convolutional variational autoencoder (CVAE) for latent space representation, a long short-term memory (LSTM) network for temporal dynamics, a dense neural network (DNN) for parameter estimation, and a classifier-pruned Bayesian optimizer (C-BO) to adaptively search and filter the latent space for optimal solutions. CBOL-Tuner demonstrates superior performance in identifying multiple optimal settings and outperforms alternative global optimization methods.

As particle accelerators grow in complexity, traditional control methods face increasing challenges in achieving optimal performance. This paper envisions a paradigm shift: a decentralized multi-agent framework for accelerator control, powered by Large Language Models (LLMs) and distributed among autonomous agents. We present a proposition of a self-improving decentralized system where intelligent agents handle high-level tasks and communication and each agent is specialized control individual accelerator components. This approach raises some questions: What are the future applications of AI in particle accelerators? How can we implement an autonomous complex system such as a particle accelerator where agents gradually improve through experience and human feedback? What are the implications of integrating a human-in-the-loop component for labeling operational data and providing expert guidance? We show two examples, where we demonstrate viability of such architecture.

Beam steering involves the calibration of the angle and position at which a particle accelerator's electron beam is incident upon the x-ray target with respect to the rotation axis of the collimator. Beam Steering is an essential task for light sources. The Linac To Undulator is very difficult to steer and aim due to the changes of each use of the accelerator there must be re-calibration of magnets. However with each use of the Beamline its current method of steering runs into issues when faced with calibrating angles and positions. Human operators spend a substantial amount of time and resources on the task. We developed multiple different feed-forward-neural networks with varying hyper-parameters, inputs, and outputs, seeking to compare their performance. Specifically, our smaller models with 33 inputs and 13 outputs outperformed the larger models with 73 inputs and 50 outputs. We propose the following explanations for this lack of performance in larger models. First, a lack of training time and computational power limited the ability of our models to mature. Given more time, our models would outperform SVD. Second, when the input size of the model increases the noise increases as well. In this case more inputs corresponded to a greater length upon the LINAC accelerator. Less specific and larger models that seek to make more predictions will inherently perform worse than SVD.

Autonomous tuning of particle accelerators is an active and challenging field of research with the goal of enabling novel accelerator technologies cutting-edge high-impact applications, such as physics discovery, cancer research and material sciences. A key challenge with autonomous accelerator tuning remains that the most capable algorithms require an expert in optimisation, machine learning or a similar field to implement the algorithm for every new tuning task. In this work, we propose the use of large language models (LLMs) to tune particle accelerators. We demonstrate on a proof-of-principle example the ability of LLMs to successfully and autonomously tune a particle accelerator subsystem based on nothing more than a natural language prompt from the operator, and compare the performance of our LLM-based solution to state-of-the-art optimisation algorithms, such as Bayesian optimisation (BO) and reinforcement learning-trained optimisation (RLO). In doing so, we also show how LLMs can perform numerical optimisation of a highly non-linear real-world objective function. Ultimately, this work represents yet another complex task that LLMs are capable of solving and promises to help accelerate the deployment of autonomous tuning algorithms to the day-to-day operations of particle accelerators.

Particle accelerators are complex machines that consist of a large number of subsystems. Although many processes are automated and feedback systems are in place, experiments and machine supervision need to be performed by a group of operators. These operators usually have an accelerator physics background and mostly know how the technology works. They especially know how to setup and tune the machine parameters for certain working points and experiments using high-level graphical user interfaces, which are connected to low-level machine control software. Due to the complexity of the machine, some subsystems of the machine are taken care of by experts, who the operators can turn to. This work shows that it is possible to support the day-to-day operation of a complex machine like a particle accelerator using a large language model (LLM), an object-oriented high-level machine control system framework, as well as a number of interfaces to knowledge bases such as the electronic logbook. The system is able to assist the operators on many levels, e.g. by producing Python scripts, which when executed perform a task defined by an input prompt to the LLM. To this end, the reasoning and action prompting paradigm (ReAct) [Yao et al., 2023] is implemented. This way a multi-expert system is realized, mimicking the real world, where the complex machine is operated by many subsystem experts.