First-hand information for everyone
This paper presents a community-informed overview and early roadmap for combining artificial intelligence (AI) with high energy physics (HEP). The work was motivated in part by discussions at the 2025 Quantum Computing and Machine Learning Workshop in Qingdao. The authors collect perspectives from researchers active in AI+HEP and aim to summarize what is happening now and what priorities might guide coordinated effort in the future.
The report reviews activity across three broad areas: experimental work, phenomenology (the link between data and theory), and theoretical physics. On the experimental side, the paper highlights China’s active role in major facilities such as the Beijing Spectrometer experiment (BESIII), the Jiangmen Underground Neutrino Observatory (JUNO), and the Large High Altitude Air Shower Observatory (LHAASO). It also notes planned projects that could benefit from AI from the design stage, including the proposed Super Tau-Charm Facility (STCF) and the Circular Electron Positron Collider (CEPC). The basic point is that modern experiments produce huge, complex datasets where AI can help find subtle signals.
The authors describe concrete AI tools already in use and under development. Convolutional neural networks and graph neural networks are applied for particle identification and track reconstruction in complex detector layouts. Generative models, for example flow-based models, are used to speed up simulation. Simpler models such as boosted decision trees and compact neural networks are used where very fast, low-latency decisions are needed, for example in trigger systems. The report also discusses deploying algorithms on hardware close to the detector — field-programmable gate arrays (FPGAs) or graphics processing units (GPUs) — and the emerging idea of co-designing detector hardware together with embedded AI algorithms.
On the theory and phenomenology side, the paper notes growing use of AI to study formal problems such as scattering amplitudes, quantum field theory calculations, and lattice simulations. Phenomenology tasks that span many experiments — event classification, unfolding (correcting measured data back to the underlying physics), parameter inference, and fast simulation — are also active areas for machine learning. The authors flag the potential role of general-purpose AI tools, including large language models and agent-based systems, for supporting research workflows and training.
The document is careful about its limits. It is an initial, partial snapshot rather than a complete community consensus. The authors report that many local efforts exist across China and East Asia — seminars, journal clubs, workshops, and internal working groups — but that these activities remain somewhat fragmented. The paper calls for more coordination, shared infrastructure, and workforce development to realize the opportunities. It also notes technical challenges that remain, such as integrating AI reliably into real-time hardware, ensuring interpretability and robustness of models, and agreeing common standards for broader community use.