Machine Learning for Invisible Dark Boson Searches at the Electron-Ion Collider

Rojae Mighty; Ankush Reddy Kanuganti

hep-phSep 18, 2026

Machine Learning for Invisible Dark Boson Searches at the Electron-Ion Collider

Authors: Rojae Mighty, Ankush Reddy Kanuganti

Organizations: Department of Physics and Astronomy, Stony Brook University, Stony Brook, NY 11794-3800, USA · Physics Department, Brookhaven National Laboratory, Upton, NY 11973-5000, USA

Abstract

We investigate whether adding $t$ , the positive magnitude of the squared nuclear four-momentum transfer, enables boosted decision trees (BDTs) to improve invisible-dark-boson selection relative to optimized rectangular cuts at the Electron-Ion Collider. We model coherent exclusive scalar and vector production at generator level in electron-gold collisions at 18 GeV by 100 GeV per nucleon. Both methods use identical weighted samples, inputs, preselection, and optimization objectives. Using only electron information, the BDT provided no consistent advantage over optimized cuts for signal selection across 11 masses for each boson type. When both methods also use $t$ , the BDT distinguishes signal from background slightly better than optimized cuts at 10 GeV for both boson types. These results motivate further investigation of machine learning in EIC dark-boson searches through exclusive processes where $t$ can be reconstructed.

Figures & tables

Figure 1: Coherent exclusive $e(k)+\mathrm{Au}(P)\to e(k^{\prime})+\mathrm{Au}(P^{\prime})+\phi(p_{\phi})$ production. The dark boson is emitted before panel (a) or after panel (b) virtual-photon exchange with intact Au. The nuclear vertex is $ZF_{A}(t)$ , with $t=-(P^{\prime}-P)^{2}$ , and the amplitude coherently sums both diagrams. The dashed line denotes a scalar or vector $\phi$ .

Figure 2: Coherent vector and scalar signal cross sections as functions of boson mass for $g_{\phi e}$ = $10^{-4}$ at $18~\mathrm{GeV}\times 100~\mathrm{GeV}$ per nucleon. Green and purple show our inclusive and selected calculations, respectively. Solid lines denote vector bosons and dashed lines denote scalar bosons. The selected curves apply the mass-dependent electron preselection described in Section II.3 . Both axes are logarithmic.

Signal	Inclusive [pb]	Selected [pb]
Vector	13.5868	0.157512
Scalar	5.96175	0.0372382

Table 1: Inclusive and selected signal cross sections from our weighted numerical integration at

m_{\phi}

= 1 GeV and

g_{\phi e}

10^{-4}

. The selected rates include the mass-dependent electron preselection.

Figure 3: Inclusive recoil-electron distributions for the public ancillary photon sample (black) and our photon proxy (blue), normalized separately within the displayed ranges. Annotations give the total-variation distances and the fractions of proxy weight outside those ranges.

Figure 4: Weighted post-Table-I distributions in $\log_{10}$ ( $t$ / $\mathrm{GeV}^{2}$ ) and $p_{T,e}$ . Rows correspond to $m_{\phi}$ = 1 and 10 GeV; columns show the coherent-photon proxy, vector signal, and scalar signal. Each panel is normalized to its total weighted class probability, so color represents shape rather than absolute cross section. Signal events use their integration weights, while photon events additionally include the stated photon-miss probability. Identical axes, bins, and color scales are used within each mass row.

Figure 5: Vector mass scan. The logarithmic horizontal axis is the vector-boson mass. The vertical axis is the BDT coupling-threshold advantage over same-input optimized rectangular cuts; zero denotes equal performance and positive values favor the BDT. Solid black circles use $Q^{2}$ , $p_{T,e}$ , $\eta_{e}$ , and $E_{e}$ . Solid blue circles give both algorithms the same four variables plus $t$ . Error bars are paired 95% event-bootstrap intervals. The dotted horizontal line marks the predefined +1% practical threshold.

Figure 6: Scalar mass scan. Axes and statistical conventions match Figure 5 . Solid black circles use the four recoil-electron inputs. Solid blue circles add $t$ to both the BDT and optimized rectangular selector.

Figure 7: Weighted post-Table-I signal efficiency $\epsilon_{S}$ versus coherent-photon-proxy efficiency $\epsilon_{\gamma}$ on the sealed 10 GeV test samples. The left and right panels show vector and scalar signals. Solid curves vary the score threshold of the unchanged trained BDTs; filled circles show the optimized rectangular operating points chosen using the validation sample. Black denotes $Q^{2}$ , $p_{T,e}$ , $\eta_{e}$ , and $E_{e}$ , while blue adds $t$ to both methods. DIS is absent from this diagnostic because no event-level DIS sample is classified.

Figure 8: TMVA BDT variable importance for the validation-selected $t$ -inclusive models across ten independent 10 GeV generation-and-training replicas. Green and purple denote vector and scalar signals. Small points show individual replicas; large circles and error bars show the mean and sample standard deviation. Importance is normalized within each fitted model.

Figure 9: Ten-replica electron-only robustness comparison. Each row is one prespecified signal benchmark. The horizontal axis is the BDT coupling-threshold advantage over same-input optimized rectangular cuts; zero means equal performance. Each point is the hierarchical median over ten complete generation-and-training replicas, and each horizontal bar is the corresponding 95% interval. All four intervals lie far inside the predefined -1% to +1% practical-equivalence region.

Signal	$m_{\phi}$ [GeV]	BDT coupling advantage [95%]	BDT wins	Decision
Vector	1.000	-0.009% [-0.029%, +0.004%]	4/10	Practical equivalence
Vector	10.000	-0.057% [-0.142%, +0.021%]	2/10	Practical equivalence
Scalar	1.000	+0.031% [-0.083%, +0.132%]	8/10	Practical equivalence
Scalar	6.310	-0.088% [-0.182%, -0.005%]	0/10	Cuts slightly favored

Table 2: Independent-seed results. All four ordinary 95% intervals lie inside the preregistered -1% to +1% practical-equivalence range; none passes the familywise superiority test.

Figure 10: Ten-replica $t$ -inclusive robustness comparison at 10 GeV. The horizontal axis is the BDT coupling-threshold advantage over same-input optimized cuts. Blue circles are hierarchical medians and horizontal bars are 95% intervals over complete independent generation-and-training replicas. The dotted vertical line marks the predefined +1% practical threshold. Both methods receive $Q^{2}$ , $p_{T,e}$ , $\eta_{e}$ , $E_{e}$ , and $t$ .

Signal	$m_{\phi}$ [GeV]	BDT advantage [95%]	BDT wins	Familywise lower	Decision
Vector	10.000	+1.208% [+1.081%, +1.326%]	10/10	+1.081%	Practical BDT superiority
Scalar	10.000	+1.361% [+1.206%, +1.529%]	10/10	+1.206%	Practical BDT superiority

Table 3: Independent-seed

t

-inclusive results. Hierarchical paired intervals combine complete generator-and-training replicas with paired event resampling. Both familywise lower bounds exceed the predefined +1% practical threshold.

Explore similar work

Searching for BSM Experimental Signatures with Large Lagrangian Models
Sep 29, 2026Ibrahim Elsharkawy, Victoria Knapp-Perez, Wahid Bhimji +1AI for Science High-Energy Physics
Department of Physics, University of Toronto and Vector Institute, Toronto, ON, Canada · NERSC, Lawrence Berkeley National Laboratory, Berkeley, California, USA · Department of Physics and Astronomy, University of California, Irvine, CA 92697 +3
Learning to bin: differentiable and Bayesian optimization for multi-dimensional discriminants in high-energy physics
Jan 12, 2026Johannes Erdmann, Nitish Kumar Kasaraguppe, Florian MausolfHigh-Energy Physics Differentiable Optimization
RWTH Aachen University, III. Physikalisches Institut A, Aachen, Germany
Hadronic Mono-Z Dark Matter Sensitivity with Flow Matching on CMS Open Data
Aug 21, 2026Hitesh Rasineni, Bhavishya ChebroluHigh-Energy Physics
VIT-AP University, Amaravati, India · VIT-AP University, Amaravati, 522241, India · Mohan Babu University, Tirupati, India +1

Sep 29, 2026J/K move · Enter open · S save

Searching for BSM Experimental Signatures with Large Lagrangian Models

Ibrahim Elsharkawy, Victoria Knapp-Perez, Wahid Bhimji, Aishik Ghosh

Department of Physics, University of Toronto and Vector Institute, Toronto, ON, Canada · NERSC, Lawrence Berkeley National Laboratory, Berkeley, California, USA · Department of Physics and Astronomy, University of California, Irvine, CA 92697+3

AI for Science High-Energy Physics Large Language Model-Guided Scientific Discovery

Open paper AskPDF arXiv

The search for physics Beyond the Standard Model (BSM) is generally limited not by the supply of theory descriptions but by the lack of discriminating experimental observations. A case in point is dark matter, where the overwhelming gravitational evidence only goes so far in distinguishing between models within a vast theory space. Exploring the space of testable model signatures may help identify overlooked experimental observables and indicate the utility of future experiments. A challenge is designing a search through model signatures outside what is found in the literature. Our primary contribution is hAIthem, a framework that combines the self-guided exploration of reinforcement learning (RL) with the broad literature-derived knowledge of LLMs. We build an RL agent that learns to find which portions of a theory's high-dimensional parameter space are not excluded under some subset of constraints by playing a Battleship-style "game" against a suite of phenomenology tools. The agent is built as a Large Lagrangian Model (LLaM), an autoregressive transformer that reads a tokenized Lagrangian, is pretrained at scale (here on ~1 billion tokens from ~10,000 Lagrangians), and is fine-tuned in a live environment. The framework then constructs a decision tree that separates RL-found regions using observables computed with established tools, and passes the remaining degenerate regions to a set of LLM agents that compete to produce realistic signatures. In this proof of concept, RL-search outperforms an evolutionary-algorithm baseline, finding more viable regions with greater physical diversity. In a restricted space of single dark scalar multiplet models, we find that hAIthem proposes interesting combinations of previously studied observables, such as the application of a halo-independent kinematic ratio to paleo-detectors.

Abstract

Explore similar work

Searching for BSM Experimental Signatures with Large Lagrangian Models

Learning to bin: differentiable and Bayesian optimization for multi-dimensional discriminants in high-energy physics

Hadronic Mono-Z Dark Matter Sensitivity with Flow Matching on CMS Open Data