⚛️ phenomenology

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

This study investigates the potential of boosted decision trees to improve invisible dark boson selection at the Electron-Ion Collider compared to optimized rectangular cuts, finding that while machine learning offers no consistent advantage using only electron information, it provides a slight improvement when nuclear four-momentum transfer ( $t$ ) is included.

Original authors: Rojae Mighty, Ankush Reddy Kanuganti

Published 2026-09-22

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Rojae Mighty, Ankush Reddy Kanuganti

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Deep within the subatomic world, particles interact in ways that sometimes leave no trace, vanishing into a realm of "dark" matter that standard physics cannot explain. Scientists at the upcoming Electron-Ion Collider, a massive machine designed to smash electrons into heavy gold nuclei, hope to catch a glimpse of these invisible particles. The idea is that if a light, hidden particle is created during a collision, it would carry away energy without hitting any detectors. Researchers would have to infer its existence by carefully measuring the recoil of the visible particles left behind, specifically the electron that bounces off and the gold nucleus that stays intact. The challenge is distinguishing this rare, mysterious signal from the overwhelming flood of ordinary particle interactions that look very similar. To do this, scientists rely on mathematical tools to sort the data, deciding which events are worth keeping and which are just background noise.

A team of researchers recently asked a specific question about how best to perform this sorting. They wanted to know if advanced computer algorithms, known as boosted decision trees, could do a better job than traditional, simpler methods of cutting out the noise. These traditional methods work like a series of rectangular boxes, accepting only events that fall within specific ranges of speed and angle. The advanced algorithms, by contrast, can draw complex, curved boundaries to separate the signal from the background. The researchers also wondered if adding one specific piece of information—the momentum transferred to the gold nucleus—would help these tools work better. In the collision, if the invisible particle takes away energy, the gold nucleus must recoil to balance the books. Measuring this recoil gives a clearer picture of what happened, but it is difficult to reconstruct in a real experiment.

To test these ideas, the team ran detailed computer simulations of collisions between electrons and gold nuclei at the energies planned for the collider. They created millions of simulated events, some containing the invisible dark boson they were looking for, and others containing only the common background noise. They then applied two different selection methods to this data. The first method used only information from the electron, such as its energy and direction. The second method added the momentum transfer to the gold nucleus as an extra clue. They compared how well the complex computer algorithm performed against the optimized rectangular boxes in both scenarios.

When the researchers looked at the results using only the electron's information, the complex algorithm offered no real advantage. Across a wide range of possible masses for the dark particle, the sophisticated machine learning tool performed essentially the same as the simpler, optimized rectangular cuts. The computer algorithm was not able to find a hidden pattern in the electron data that the simpler method missed. This suggests that for this specific type of search, if you cannot measure the nucleus, the extra complexity of the machine learning does not pay off. The simpler method is just as effective at finding the signal.

However, the story changed when the researchers included the momentum transfer of the gold nucleus in the analysis. In this scenario, the complex algorithm did show a small but consistent improvement over the rectangular cuts, but only for heavier dark particles. At a mass of 10 GeV, the algorithm was able to distinguish the signal from the background slightly better than the simple method. The improvement was modest, amounting to a roughly one percent gain in the ability to detect the particle, but it was a repeatable result that held up across many different simulations. The researchers found that the algorithm used the extra information about the nucleus to make finer distinctions that the simple rectangular boxes could not capture.

The study concludes that while machine learning is not a magic bullet for every search, it can provide a tangible benefit when specific, hard-to-measure information is available. The key finding is that the advantage of using complex algorithms depends entirely on whether the experiment can reconstruct the recoil of the nucleus. If the nucleus is measured accurately, the algorithm can leverage that data to find a slight edge. If the nucleus is not measured, the simpler methods remain just as good. This result guides the next steps for the Electron-Ion Collider: the focus now shifts to building detectors that can accurately measure the recoiling gold nucleus. If they can do that, the team believes the small advantage offered by machine learning could become a real tool in the search for invisible dark matter.

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

Problem Statement
The Electron-Ion Collider (EIC) offers a unique opportunity to search for light, weakly coupled particles beyond the Standard Model, specifically invisibly decaying scalar and vector dark bosons produced via coherent exclusive processes ( $e + \text{Au} \to e + \text{Au} + \phi$ ). In these events, the gold nucleus remains intact, and the dark boson is invisible, requiring inference from the recoil electron. Previous analyses relied on "optimized rectangular cuts" applied to four electron kinematic variables: momentum transfer squared ( $Q^2$ ), transverse momentum ( $p_{T,e}$ ), pseudorapidity ( $\eta_e$ ), and energy ( $E_e$ ).

The central question addressed in this work is whether Machine Learning (ML), specifically Boosted Decision Trees (BDTs), can outperform these optimized rectangular cuts in signal selection. Furthermore, the study investigates whether including the nuclear momentum transfer squared ( $t$ )—a variable reconstructible from the intact outgoing ion but not fully determined by the electron alone when $\phi$ is invisible—provides a distinct advantage to ML methods over simple cut-based approaches.

Methodology
The authors conducted a controlled, fair comparison between BDTs and a strong multistart rectangular selector using identical weighted samples, inputs, preselections, and optimization objectives.

Signal and Background Modeling:
- Signal: Coherent exclusive production of scalar and vector bosons ( $m_\phi = 0.01$ –$10$ GeV) in $18 $GeV$ \times$ $100$ GeV/nucleon electron-gold collisions. Events were generated via direct numerical integration of three-body phase space and matrix elements, including Helm form factors and interference terms.
- Background: The dominant background is coherent photon bremsstrahlung ( $e + \text{Au} \to e + \text{Au} + \gamma$ ) where the photon is missed. The authors generated a reproducible photon proxy using the massless-vector limit of their coherent matrix element, scaled by a miss probability of $10^{-6}$ . Deep-inelastic scattering (DIS) was treated as an irreducible background with a fixed effective rate (95% veto survival), not classified at the event level.
Input Sets:
1. Electron-only: $\log_{10}(Q^2)$ , $p_{T,e}$ , $\eta_e$ , and $E_e$ .
2. Extended: The four electron variables plus the nuclear momentum transfer $t$ .
Optimization: Both methods acted as a second-stage refinement after a mass-dependent electron preselection. They maximized an approximate background-dominated significance ratio ( $R_Z$ ).
Validation and Robustness:
- Data was split into training, validation, and independent test sets to prevent leakage.
- A "practical equivalence" threshold was defined: a BDT advantage of $>1\%$ in coupling threshold (corresponding to $\approx 2.03\%$ improvement in $R_Z$ ) was required to claim a useful ML improvement.
- Robustness was tested via ten independent generation-and-training replicas at specific mass benchmarks (1, 6.31, and 10 GeV) using hierarchical bootstrapping to account for seed and sample variation.

Key Contributions

Fair ML Benchmarking: The study establishes a rigorous protocol for comparing ML against optimized cuts in high-energy physics, ensuring that any performance gain is attributed to the decision boundary complexity rather than baseline selection differences.
Role of Nuclear Recoil ( $t$ ): The paper isolates the specific contribution of the nuclear momentum transfer variable ( $t$ ) to selection performance, distinguishing between information available from the electron alone versus the full exclusive system.
Systematic Robustness: By employing independent seed replicas and hierarchical bootstrapping, the authors demonstrate that their conclusions regarding the equivalence or superiority of methods are stable against statistical fluctuations in event generation and model training.

Results

Electron-Only Inputs: Across 11 masses for both scalar and vector bosons, the BDT provided no consistent advantage over optimized rectangular cuts. The coupling-threshold differences fluctuated around zero, with no point passing the familywise-controlled lower bound or reaching the $+1\%$ practical threshold. The optimized rectangular selector captured essentially all the improvement possible with electron-only information.
Extended Inputs (Including $t$ ):
- At low masses, the addition of $t$ yielded negligible differences between methods.
- At 10 GeV, a modest but reproducible advantage for the BDT emerged.
- Vector Signal (10 GeV): BDT advantage of $+1.208\%$ (95% interval: $[1.081\%, 1.326\%]$ ).
- Scalar Signal (10 GeV): BDT advantage of $+1.361\%$ (95% interval: $[1.206\%, 1.529\%]$ ).
- In the robustness study, the BDT won all 10 independent replicas for both signal types at 10 GeV. The familywise lower bounds exceeded the predefined $+1\%$ threshold.
Variable Importance: In the 10 GeV models, $t$ was ranked as the most important variable by TMVA, with mean importances of $\approx 0.28$ (vector) and $\approx 0.29$ (scalar), significantly higher than the electron variables.

Significance and Claims
The paper claims that while ML offers no practical benefit over optimized cuts when relying solely on electron kinematics, the inclusion of the nuclear recoil variable $t$ enables a small, repeatable advantage for BDTs at high boson masses (specifically 10 GeV).

The authors emphasize that this result is modest:

The improvement corresponds to a $\approx 2.5\%$ – $2.8\%$ increase in the significance proxy ( $R_Z$ ) relative to optimized cuts.
The advantage is attributed to the BDT's ability to model nonlinear correlations between $t$ and electron variables, which rectangular cuts cannot capture.
The study does not claim that this advantage is guaranteed in a real detector environment. The current analysis uses generator-level truth values for $t$ and ignores detector effects (acceptance, resolution, smearing).

Conclusion and Next Steps
The primary conclusion is that the information contained in $t$ is valuable for dark boson searches and that ML can exploit this information slightly better than simple cuts at high masses. However, the authors state that the next necessary step is a full detector simulation to determine if the reconstructed $t$ retains enough resolution and correlation structure for the BDT advantage to survive realistic measurement effects. The paper serves as a motivation for such future studies rather than a definitive proof of ML superiority in the final experimental analysis.

More like this