Agentic AI -- Physicist Collaboration in Experimental Particle Physics: A Proof-of-Concept Measurement with LEP Open Data

The gist The analysis presents an AI agentic measurement of thrust distribution in e+e− collisions at √s = 91.2 GeV using archived ALEPH data, suggesting that precision physics leveraging open LEP data provides an ideal testing ground for developing advanced AI systems for scientific applications

Context and Motivation

Achieving Standard Model predictions at the level of precision required for future collider programs is essential to ensure robust control of systematic uncertainties in the interpretation of potential new physics anomalies Precision QCD studies can be pursued at hadron colliders such as the LHC, but they are most cleanly and systematically tested in lepton collider environments The e+e− alliance has successfully curated and reanalyzed archived datasets from the Large Electron–Positron collider (LEP), opening new opportunities enabled by three decades of theoretical advances and modern experimental techniques This combination provides a uniquely powerful environment for high-precision studies The successful curation and use of archived open data establishes the conditions for a theory–experiment loop in which AI systems can propose new calculations and directly confront them with high-precision data

AI Agent Usage in Analysis

The analysis utilizes an agentic workflow where the physicist formulates goals and provides expert direction through iterative prompts, while the agent executes the corresponding implementation steps The operating model is human-inthe-loop, where the agent is used to execute and monitor multi-step workflows, enforce consistent plotting conventions, and connect analysis statements in the note to concrete script paths The agent was supplied with context comprising reference materials like published ALEPH thrust paper [12], prior ROOT-based analysis macros, internal notes documenting the miniDST format, and LEP open-data files No pre-written unfolding or systematic-uncertainty code was provided; instead, the agent constructed these from the physics descriptions in the prompts

Analysis Method and Workflow

The core analysis involves Iterative Bayesian Unfolding (IBU) implemented in RooUnfold, with a nominal run of Niter = 5 and applying post-unfolding hadronic event correction from tgenBefore/tgen The response filling uses event-by-event matching through uniqueID to define the selected detector-level phase space The analysis incorporates several systematic variations, including iteration variations (iter4, iter6), experimental variations (ntpc7, pt04, ech10), and theory reweighting branches (pythia8, herwig, sherpa) The total bin uncertainty is formed by combining all six components in quadrature: statistical, unfolding iteration, experimental detector, theory modeling, hadronic event-selection correction, and ISR correction

Results and Uncertainty Decomposition

The final unfolded thrust distribution shows good agreement with the published ALEPH 2004 result (χ2/ndf = 0.36 over 42 bins), with the largest discrepancy at the T → 1 dijet endpoint where MC modeling uncertainties are largest The total bin-to-bin correlation matrix shows that the theory component has the strongest and most extended off-diagonal structure, dominating the total correlation matrix The dominant source across the central measurement range is theory, with experimental envelope typically being 5–15% and largest at T ≲ 0.62 where the steeply rising spectrum amplifies small absolute shifts

Conclusion and Future Work

The precision QCD motivation for this work is the persistent e+e− αs tension, as recent thrust-based extractions yield αs(mZ) ≃ 0.1136, approximately 3.5σ below the world average [20] The covariance and correlation products from this baseline are designed as direct inputs to precision αs fits This establishes a first end-to-end repository baseline for thrust with IBU and defines the template for extension to C-parameter and heavy-jet mass, including future multidimensional correlation studies The AI-agent operating model demonstrated in this cycle provides a scalable framework for precision QCD as a testing ground for closed theory–experiment loops in fundamental physics

Challenges and Scaling

The most significant open challenge is endpoint-bin instability, where the final bin (0.99 < T < 1.00) is sensitive to MC modeling, hadronic correction extrapolation, and normalization convention differences The key unsolved challenge is encoding implicit physics priors as testable assertions through formal specification languages or structured physicist–agent dialogue protocols The current mitigation relies on structured sign-off points, automated integrity checks, and version-controlled provenance The key is to develop methods to translate expert knowledge into machine-executable tests for robust AIassisted experimental physics

Workflow and Reproducibility

The executable workflow map uses two entrypoints, run thrust prelim workflow.py for nominal and systematic reruns, and build prelim uncertainty summary.py for uncertainty assembly The reproduction contract explicitly documents the command sequence for running the full 1994 chain with all configured reruns The workflow pseudocode summarizes the logic for computing signed branch shifts, building component covariances, and combining them into a total covariance matrix

Detector Performance Validation

Particle-category kinematic validation compares data and MC reco distributions for each pwflag category to ensure detector modeling of inputs is uniform across categories Agreement in the charged-track category anchors the full detector response modeling The comparison with modern generators confirms that lepton-category partitioning does not induce a hidden detector asymmetry in the thrust inputs Agreement in this panel supports a controlled transfer of neutral energy flow into the unfolded thrust response

Systematic Combination

The total uncertainty combines all six components in quadrature, with the theory component dominating the off-diagonal structure The final result is presented using a fit-window covariance product for αs extractions to avoid regions of large uncertainty and poor MC-model control near kinematic boundaries This establishes a template for extension to additional event-shape observables like the C-parameter The AI-agent operating model demonstrated in this cycle provides a scalable framework for precision QCD as a testing ground for closed theory–experiment loops in fundamental physics The analysis propagates six independent uncertainty components, with the total uncertainty combining all six sources to form the final covariance matrix suitable for precision phenomenology

Final Comparison

The final unfolded thrust distribution shows agreement across the bulk of the spectrum (0.62 ≤ T < 0.99), with a largest discrepancy at the endpoint (0.99 < T < 1.00, ratio 1.30±0.27) This is physically expected as the dijet limit is sensitive to MC modeling of soft hadrons The total correlation matrix shows strong positive correlations in the high-thrust region where theory and hadronic-correction components dominate their shifts This full covariance matrix must be used in any downstream fit or comparison

Acknowledgments

The authors would like to thank Roberto Tenchini and Guenther Dissertori from the ALEPH Collaboration for their useful comments and suggestions on the use of ALEPH data This work has been supported by Schmidt Sciences Foundation (to A.B.) and Department of Energy, Office of Science, under Grant No. DE-SC0011088 (to Y.-J.L.)

References

The references cited support the analysis and the broader context of AI in physics research The paper cites recent case studies in AI+scientist and theory-assistant case studies to expand its motivation The work is grounded in established precision measurements from the ALEPH collaboration This paper cites recent case studies in AI+scientist and theory-assistant case studies to expand its motivation<ref:53 C.

1. Automated assertion generation for physics constraints: The AI system can be improved by developing "physics-specific assertions — such as verifying that the unfolded distribution integrates to unity, that the response matrix is properly normalized, or that refolded spectra agree with detector-level data — could be encoded as unit tests" to run after each agent iteration.

2. Formalization of expert intuition into machine-executable constraints: The system can be enhanced by "Developing methods to translate such expert knowledge into machine-executable tests — whether through formal specification languages, learned anomaly detectors, or structured physicist–agent dialogue protocols — is an important direction for making AI-assisted experimental physics robust and scalable."

3. Automated tracking of systematic inter-bin correlations: The system can be improved by implementing a mechanism to "systematic inter-bin correlations use the statistical correlation matrix as a template (Cov(c)ij = ∆c(i) ∆c(j) ρstatij); this is an approximation that assumes systematic and statistical correlations share the same bin-to-bin structure."

4. Dynamic uncertainty decomposition reporting: The system can be enhanced by moving beyond envelope-only summaries to explicitly report component behavior, as the workflow has migrated from envelope-only views to explicit component-level plots, including signed shifts, grouped uncertainty structure for better source resolution.

5. Automated validation of model dependencies: The system can be improved by implementing automated regression tests that verify that modifying one systematic variation does not silently shift the unfolded spectrum outside a predefined tolerance, addressing the risk where the theory envelope is the dominant uncertainty contribution across the central measurement range.

We present an agentic measurement of the thrust distribution in e+e- collisions at sqrt s =91.2 GeV using ALEPH archived data. This work was carried out through an AI agent-physicist collaboration, using interactive natural-language prompting of Anthropic Claude and OpenAI Codex. Under physicist guidance, the agents wrote and executed all analysis code and drafted the majority of the text included in this paper. A fully-corrected thrust spectrum is obtained via Iterative Bayesian Unfolding and Monte Carlo based corrections. The principal contribution of this work is the demonstration of the agentic analysis, with thrust serving as a benchmark observable for the method. Dedicated precision measurements of thrust are reported in the community elsewhere. This work represents a step toward a theory-experiment loop in which AI agents assist with experimental measurements and theoretical calculations, and synthesize insights by comparing the results, thereby accelerating the cycle that drives discovery in fundamental physics. Our work suggests that precision physics, leveraging the open LEP data and advanced theoretical landscape, provides an ideal testing ground for developing AI agent systems for scientific applications.