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

Anthony Badea; Yi Chen; Marcello Maggi; Yen-Jie Lee; Electron-Positron Alliance

hep-exMar 5, 2026

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

Authors: Anthony Badea, Yi Chen, Marcello Maggi, Yen-Jie Lee, Electron-Positron Alliance

Organizations: University of Chicago, Enrico Fermi Institute, 60637 IL, USA · Vanderbilt University, Nashville, Tennessee, USA · Istituto Nazionale di Fisica Nucleare, Bari Division, BA 70126, Italy · Laboratory for Nuclear Science, Massachusetts Institute of Technology, Cambridge, MA 02139, USA

Abstract

We present an AI agentic measurement of the thrust distribution in $e^{+}e^{-}$ collisions at $\sqrt{s}=91.2$ ~GeV using archived ALEPH data. The analysis and all note writing is carried out entirely by AI agents (OpenAI Codex and Anthropic Claude) under expert physicist direction. A fully corrected spectrum is obtained via Iterative Bayesian Unfolding and Monte Carlo based corrections. 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 advanced AI systems for scientific applications.

Figures & tables

Sample	Total entries	Selected entries
1994 data reco	1,365,440	1,326,312
1994 MC reco	771,597	751,861
1994 MC gen ( tgen )	771,597	771,597
1994 MC gen-before ( tgenBefore )	973,769	973,769

Table 1: Nominal sample bookkeeping for the thrust unfolding chain. Selected entries correspond to passEventSelection =1. For generator-level trees, all entries are retained by construction.

pwflag	Category
0	charged tracks
1	charged leptons (tracker + ECAL category)
2	charged leptons (tracker + muon category)
3	$V^{0}$ candidates
4	photons
5	neutral hadrons

Table 2: Particle-category labels used in archived ALEPH reconstructed events.

Selection item	Requirement
Charged category	pwflag $\in[0,2]$
Charged acceptance	$\|\cos\theta\|\leq 0.94$
Charged transverse momentum	$p_{\mathrm{T}}\geq 0.2$ GeV
Charged impact parameter	$\|d_{0}\|\leq 2$ cm
Charged longitudinal impact	$\|z_{0}\|\leq 10$ cm
Charged hit quality	$N_{\mathrm{TPC}}\geq 4$

Table 3: Object-level selections used to build the nominal thrust input particle set, with one line per requirement.

Selection item	Requirement
Ntuple preselection	passesNTupleAfterCut = 1
Total charged energy	$E_{\mathrm{trk}}^{\mathrm{tot}}\geq 15$ GeV
Sphericity-axis acceptance	$\|\cos\theta_{\mathrm{Sph}}\|\leq 0.82$
Charged multiplicity	$N_{\mathrm{trk}}\geq 5$
Total multiplicity	$N_{\mathrm{trk}}+N_{\mathrm{neu}}\geq 13$
Visible energy guard	$E_{\mathrm{vis}}\geq 0$ GeV (nominally inactive)

Table 4: Nominal event-level hadronic selection requirements, including guards present in code but inactive in the nominal configuration.

Panel of Figure 1 — Figure 1: Complete pwflag =0 kinematic set.

Panel of Figure 2 — Figure 2: Complete pwflag =1 kinematic set.

Panel of Figure 3 — Figure 3: Complete pwflag =2 kinematic set.

Panel of Figure 4 — Figure 4: Complete pwflag =3 kinematic set.

Panel of Figure 5 — Figure 5: Complete pwflag =4 kinematic set.

Panel of Figure 6 — Figure 6: Complete pwflag =5 kinematic set.

Figure 7: Left:

\cos\theta

for pwflag=5 (neutral hadrons), data vs MC with event selection applied; the yellow band marks the cleaning cut region

-0.19\leq\cos\theta<-0.18

. Right: data with vs without event selection, showing the suppression introduced by the cut.

Figure 8:

\phi

distribution for pwflag=4 photons (left) and pwflag=5 neutral hadrons (right), comparing data reco, MC reco, and MC gen. Dashed vertical lines mark the 24-fold HCAL module boundaries. The modulation present in MC reco and gen but absent in data at the same phase points to a generator-level origin.

Panel of Figure 9 — Figure 9: Event-level observables used in the nominal hadronic selection; legends include detector-level data/MC and generator-level references.

Item	Nominal configuration
Observable	Thrust $T$
Range	$0.5\leq T<1.0$
Binning	50 uniform bins
Reco/gen matching key	uniqueID
Reco selection branch	passEventSelection
Nominal IBU iterations	5

Table 5: Nominal analysis configuration used in the thrust Iterative Bayesian Unfolding workflow.

Figure 10: Left: reco efficiency (matched/all gen) as a function of

T_{\mathrm{gen}}

. Right: thrust distributions for all gen, matched, and miss events (normalized to unit area). Miss events are concentrated at low thrust and low multiplicity, consistent with events failing the hadronic event selection.

Figure 11: Distribution of

(T_{\mathrm{reco}}-T_{\mathrm{gen}})

for matched MC event pairs, integrated over all

T_{\mathrm{gen}}

. The distribution is narrow and centered at zero, confirming no significant systematic bias in the thrust reconstruction.

Component	Inputs	Combination rule
Iteration	iter4 , iter6	Envelope per bin: $\max\|\Delta_{\mathrm{iter}}\|$
Track/event	ntpc7 , pt04 , ech10 , with_met	Quadrature of absolute shifts
Neutral	nes_up , nes_down , ner	NES envelope, then quadrature with NER
Theory	pythia8 , herwig , sherpa	Envelope across generators
Hadronic corr.	tgen / tgenBefore ratio	Propagated bin-by-bin
Exp. total	track/event $+$ neutral	$\Delta_{\mathrm{exp}}=\sqrt{\Delta_{\mathrm{t/e}}^{2}+\Delta_{\mathrm{neu}}^{2}}$

Table 6: Systematics combination scheme for the thrust baseline. Each row gives the input systematic variations and the combination rule used to form the per-bin uncertainty contribution. The experimental subtotal (Exp. total) and final total are formed by quadrature of their respective sub-components. The cross-check branch ( no_neutrals ) is used for stress-testing only.

Figure 12: Nominal response matrices for thrust: response counts (left) and response probability (right).

Figure 13: Nominal unfolded covariance matrix (left) and corresponding correlation matrix (right).

Figure 14: Signed per-bin shifts for the statistical (left) and unfolding-iteration (right) uncertainty components. The statistical component is nearly flat and small across the spectrum. The iteration panel shows

N_{\mathrm{iter}}=4

and

6

relative to the nominal

N_{\mathrm{iter}}=5

; the envelope (dashed black) is sub-percent in the bulk, confirming a stable regularization plateau.

Group	Branch	What is varied	Combination
Track/event	ntpc7	TPC hit requirement $N_{\mathrm{TPC}}>7$ (nominal $>4$ )	Quadrature of four shifts
	pt04	Charged $p_{T}>0.4$ GeV threshold (nominal $>0.2$ GeV)
	ech10	Charged-energy cut $>10$ GeV (nominal $>15$ GeV)
	with_met	Missing- $E_{T}$ added as pseudo-particle
Neutral response	nes_up/dn	Neutral energies scaled up/down; probes ECAL energy-scale	Envelope then quadrature with ner
	ner	Neutral objects randomly removed; probes reco efficiency	Quadrature with nes envelope

Table 7: Experimental systematic variations. Track/event branches are combined in quadrature; neutral-response branches as envelope+quadrature.

\Delta_{\mathrm{exp}}

is the quadrature sum of the two sub-totals.

Figure 15: Signed per-bin shifts for the seven experimental variation branches, with the group envelope (dashed black). The four track/event branches ( ntpc7 , pt04 , ech10 , with_met ) and three neutral-response branches ( nes_up , nes_down , ner ) are shown; the

y

-axis is clipped to

\pm 50\%

Figure 16: Comparison of the charged-only (red open circles) and energy-flow (black filled circles) unfolded thrust distributions. Both are processed through the complete analysis chain: IBU (

N_{\mathrm{iter}}=5

), hadronic correction

C_{\mathrm{had}}(j)

, and ISR correction

C_{\mathrm{ISR}}(j)

, with separate response matrices for each particle-level definition. Error bars represent statistical uncertainties only. Both distributions are normalized to

(1/N)\,\mathrm{d}N/\mathrm{d}T

. The ratio panel shows charged-only / energy-flow; the dashed line marks unity. The

{\sim}23\%

central-range shift reflects the different particle-level definitions of the two observables, not a detector systematic effect, motivating the use of neutral-response variations ( nes_up , nes_down , ner ) to probe calorimeter uncertainties within the energy-flow observable.

Figure 17: Signed per-bin shifts for the three theory reweighting branches ( pythia8 , herwig , sherpa ) with the theory envelope (dashed black). Generator differences grow at lower thrust, reflecting parton-shower and hadronization modeling uncertainties. This is the dominant source across the central measurement range. The

y

-axis is clipped to

\pm 50\%

Panel of Figure 18 — Figure 18: Absolute per-bin fractional uncertainties ( $y$ -axes limited to $50\%$ ). Top left: statistical and unfolding-iteration components. Top right: seven experimental variation branches with group envelope. Bottom left: theory reweighting branches with theory envelope. Bottom right: all groups overlaid; dashed black shows total.

Figure 19: Left: final unfolded thrust distribution with total uncertainty band compared to the published ALEPH 2004 result. The ratio panel shows bin-by-bin agreement across the full measurement range, with the gray band representing the total uncertainty envelope. Right: comparison with OmniFold and prior IBU spectra; the ratio panel shows each method relative to the current unfolded nominal. Compatibility metrics are

\chi^{2}/\mathrm{ndf}=0.663

(OmniFold),

0.421

(prior IBU), and

0.361

(ALEPH 2004), evaluated over 42 bins using the diagonal of the total covariance.

Figure 20: Total bin-to-bin correlation matrix for the final result, combining statistical, unfolding-iteration, experimental, theory, hadronic-correction, and ISR-correction components using the template_correlated covariance model. The maximum off-diagonal correlation is

|\rho_{ij}|=0.889

$T$ range	$(1/\sigma)\,\mathrm{d}\sigma/\mathrm{d}T$	Stat.	Iter.	Exp.	Theory	Had. corr.	Total
		[%]	[%]	[%]	[%]	[%]	[%]
[0.55,0.56)	—
[0.56,0.57)	$6.30\times 10^{-5}$	88.8	6.9	467.8	67.7	139.7	500.9
[0.57,0.58)	0.0012	28.8	1.2	46.7	38.3	70.5	97.2
[0.58,0.59)	0.0015	14.5	0.4	25.0	59.5	49.6	82.7
[0.59,0.60)	0.0032	15.8	5.2	35.8	53.5	35.5	75.3

Table 8: Per-bin fractional uncertainties across the full measurement range (

0.55\leq T<1.0

). The second column gives the nominal unfolded

(1/\sigma)\,\mathrm{d}\sigma/\mathrm{d}T

value; all remaining columns are absolute fractional uncertainties in percent. Bins with zero nominal spectrum (dashes) are below threshold. Exp. denotes the total experimental component (track/event and neutral sub-components combined in quadrature), and Had. corr. denotes the hadronic event-selection correction uncertainty.

$T$ range	Our $(1/\sigma)\,\mathrm{d}\sigma/\mathrm{d}T$	Our total unc. [%]	ALEPH 2004 $(1/\sigma)\,\mathrm{d}\sigma/\mathrm{d}T$	ALEPH 2004 total unc. [%]
[0.55,0.56)	—	—	—	—
[0.56,0.57)	$6.30\times 10^{-5}$	500.9	—	—
[0.57,0.58)	0.0012	97.2	—	—
[0.58,0.59)	0.0015	82.7	0.0011	61.2
[0.59,0.60)	0.0032	75.3	0.0023	54.1
[0.60,0.61)	0.0042	39.3	0.0035	41.7

Table 9: Comparison of the nominal unfolded

(1/\sigma)\,\mathrm{d}\sigma/\mathrm{d}T

and total fractional uncertainty with the published ALEPH 2004 result (HEPData ins:636645, Table 54) across the full measurement range (

0.55\leq T<1.0

). The ALEPH 2004 total uncertainty combines statistical, sys 1 , and sys 2 components in quadrature. Dashes indicate bins not covered by one or both measurements.

	$(1/\sigma)\,\mathrm{d}\sigma/\mathrm{d}T$		Stat. [%]		Exp. / Sys. 1 [%]		Theory / Sys. 2 [%]		Total [%]
$T$ range	Ours	ALEPH	Ours	ALEPH	Ours	ALEPH	Ours	ALEPH	Ours	ALEPH
[0.55,0.56)	—
[0.56,0.57)	$6.30{\times}10^{-5}$	—	88.8	—	467.8	—	67.7	—	500.9	—
[0.57,0.58)	$1.22{\times}10^{-3}$	—	28.8	—	46.7	—	38.3	—	97.2	—
[0.58,0.59)	$1.48{\times}10^{-3}$	$1.15{\times}10^{-3}$	14.5	29.6	25.0	45.2	59.5	28.7	82.7	61.2
[0.59,0.60)	$3.23{\times}10^{-3}$	$2.29{\times}10^{-3}$	15.8	21.8	35.8	40.6	53.5	28.4	75.3	54.1

Table 10: Per-bin fractional uncertainty breakdown comparison between this analysis and the published ALEPH 2004 result (HEPData ins:636645, Table 54). For this analysis, Stat. is derived from the RooUnfold covariance, Exp. is the quadrature sum of all experimental track/event and neutral variations, and Theory is the max envelope of the three generator reweightings. For ALEPH 2004, Sys. 1 and Sys. 2 correspond to the two systematic components provided in the HEPData table. All values are absolute fractional uncertainties in percent; Total combines all components in quadrature. Dashes indicate bins outside the respective measurement range.

Figure 21: Per-source correlation matrices for the six independent uncertainty components: statistical (top left), unfolding-iteration (top center), experimental (top right), theory (bottom left), hadronic correction (bottom center), and ISR correction (bottom right). All matrices use the same color scale ( $-1$ to $+1$ , red/blue). The theory component shows the strongest and most extended off-diagonal structure, dominating the total correlation matrix.

Figure 22: Per-source covariance matrices for the six independent uncertainty components: statistical (top left), unfolding-iteration (top center), experimental (top right), theory (bottom left), hadronic correction (bottom center), and ISR correction (bottom right). Each panel uses a symmetric color scale centered on zero (red: positive, blue: negative). The theory component dominates the total covariance in both magnitude and off-diagonal extent.

Figure 23: Total covariance matrix, combining all six uncertainty components. The off-diagonal structure is dominated by theory correlations across the central thrust range. The full covariance matrix must be used in any downstream fit; diagonal uncertainties alone would underestimate the true statistical power of the result. The fit-window version ( $0.505\leq T<1.00$ ) is recommended for $\alpha_{s}$ extractions.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 24 — Figure 24: Schematic analysis flow for the thrust preliminary workflow. This replaces the dense workflow table and shows the execution order and branch structure used in this internal note.

Explore similar work

AI Agents Can Already Autonomously Perform Experimental High Energy Physics
Mar 20, 2026Eric A. Moreno, Samuel Bright-Thonney, Andrzej Novak +2Autonomous Research Agents AI Agents for Scientific Discovery
Department of Physics, Massachusetts Institute of Technology · NSF AI Institute for Artificial Intelligence and Fundamental Interactions · CERN
Collider-Bench: Benchmarking AI Agents with Particle Physics Analysis Reproduction
May 13, 2026Darius A. Faroughy, Sofia Palacios Schweitzer, Ian Pang +2AI Agents for Scientific Discovery AI Agent Evaluation
New High Energy Theory Center Department of Physics & Astronomy Rutgers University · Faculty of Computing & Data Sciences Boston University
Agentic Artificial Intelligence for Multistage Physics Experiments at a Large-Scale User Facility Particle Accelerator
Sep 21, 2025Thorsten Hellert, Drew Bertwistle, Simon C. Leemann +2AI for Science Agentic Workflows
Lawrence Berkeley National Laboratory, Berkeley, California 94720, USA

Mar 20, 2026J/K move · Enter open · S save

AI Agents Can Already Autonomously Perform Experimental High Energy Physics

Eric A. Moreno, Samuel Bright-Thonney, Andrzej Novak, Dolores Garcia+1

Department of Physics, Massachusetts Institute of Technology · NSF AI Institute for Artificial Intelligence and Fundamental Interactions · CERN

Autonomous Research Agents AI Agents for Scientific Discovery Data Analysis Agents LLM Agent Harnesses High-Energy Physics

Open paper AskPDF arXiv

Large language model-based AI agents are now able to autonomously execute substantial portions of a high energy physics (HEP) analysis pipeline with minimal expert-curated input. Given access to a HEP dataset, an execution framework, and a corpus of prior experimental literature, we find that Claude Code succeeds in automating all stages of a typical analysis: event selection, background estimation, uncertainty quantification, statistical inference, and paper drafting. We argue that the experimental HEP community is underestimating the current capabilities of these systems, and that most proposed agentic workflows are too narrowly scoped or scaffolded to specific analysis structures. We present a proof-of-concept framework, Just Furnish Context (JFC), that integrates autonomous analysis agents with literature-based knowledge retrieval and multi-agent review, and show that this is sufficient to plan, execute, and document a credible high energy physics analysis. We demonstrate this by conducting analyses on open data from ALEPH, DELPHI, and CMS to perform electroweak, QCD, and Higgs boson measurements. We present two of those results in a condensed short paper form -- a CMS Run1 Open Data

H\to τ^+τ^-

to demonstrate performance on a well-established result, and the first Lund plane measurement on LEP data -- a genuinely novel result and, to our knowledge, the first produced autonomously by an AI agent. Rather than replacing physicists, these tools promise to offload the repetitive technical burden of analysis code development, freeing researchers to focus on physics insight, truly novel method development, and rigorous validation. Given these developments, we advocate for new strategies for how the community trains students, organizes analysis efforts, and allocates human expertise.

Abstract

Explore similar work

AI Agents Can Already Autonomously Perform Experimental High Energy Physics

Collider-Bench: Benchmarking AI Agents with Particle Physics Analysis Reproduction

Agentic Artificial Intelligence for Multistage Physics Experiments at a Large-Scale User Facility Particle Accelerator