Higgs Signal Strength Estimation with Machine Learning under Systematic Uncertainties

arXiv:2509.00672 · hep-ph, hep-ex · Submitted 2025-08-31 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: High Energy Physics: Phenomenology. The day in particle phenomenology: models, signatures and constraints.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Higgs Signal Strength Estimation with Machine Learning under Systematic Uncertainties".

Mira: Here's the gist of the paper. Higgs Signal Strength Estimation with Machine Learning under Systematic Uncertainties presents a dedicated graph neural network (GNN)-based methodology for extracting Higgs boson signal strength µ while incorporating systematic uncertainties<ref:2509.00672H1,three branches: a deterministic GNN that processes kinematic variables unaffected by nuisance parameters, and an uncertainty-aware GNN that handles inputs modulated by systematic effects through gated attention-based message passing<ref:2509.00672H3,The resulting binned classifier outputs are used to construct a Poisson likelihood, which enables profile likelihood scans over signal strength, with nuisance parameters profiled out via numerical optimization<ref:2509.006724,This framework falls into the category of decorrelation techniques, where the goal is to suppress the sensitivity of classifier outputs to nuisance variations while preserving discriminative power between signal and background<ref:2509.006724,We validate our method using the FAIR Universe Higgs Uncertainty Challenge dataset, yielding accurate estimation of signal strength µ and its 68.27% confidence interval, achieving competitive coverage and interval widths in large-scale pseudo-experiments<ref:2509.00672#pg6>.

Kai: The architecture features a dual-branch design that separates the treatment of deterministic and uncertainty-sensitive features<ref:2509.006723,The deterministic branch is designed to process features that are unaffected by systematic variations<ref:2509.006721,The uncertainty-aware branch is designed to process node features nunc, edge features eunc, and global features gunc that are sensitive to systematic variations<ref:2509.00672#pg5>.

Mira: During training we sample nuisance-parameter configurations and aggregate the loss across them, promoting stability of the classifier under systematic shifts and effectively decorrelating its outputs from nuisance variations<ref:2509.006721,The training strategy incorporates stochastic sampling over nuisance parameters to marginalize systematic effects during training<ref:2509.00672#pg2>.

Kai: Data Representation and Graph Structure Each event is encoded as a graph with five nodes, corresponding to the final-state objects: lepton, tau-jet (τhad), leading jet (j1), subleading jet (j2), and missing transverse momentum (MET)<ref:2509.006723,For each node, we construct two types of features: deterministic and uncertainty-aware<ref:2509.006723,The deterministic node features include the lepton transverse momentum pT together with pseudorapidity η and azimuthal angle φ of all the visible final objects<ref:2509.006723,The uncertainty-aware node features consist of quantities sensitive to the three calibration-type nuisance parameters<ref:2509.00672#pg4>.

Paper summary: Kai: So we're looking at this paper now called "Higgs Signal Strength Estimation with Machine Learning under Systematic Uncertainties". It seems like the main idea is using a graph neural network to figure out the Higgs boson signal strength while dealing with these nasty systematic uncertainties that always plague experiments.

Mira: Right. Basically, they have this new way of doing things where they use a dual-branch GNN architecture. One branch handles things that are pretty fixed and don't change much because of those nuisance parameters, and the other branch specifically deals with the inputs that are sensitive to those systematic effects through some gated attention message passing

2509.00672H1, 2509.00672H3: .

Lev: That sounds like they're trying to separate the deterministic parts of the kinematics from the parts that are affected by things like detector calibrations or theoretical modeling two thousand five hundred nine point zero zero six seven two three. I wonder if separating them actually helps when you're dealing with a complex system?

Kai: Exactly. The paper claims this approach promotes stability because during training, they sample different configurations of those nuisance parameters and average the loss across them

two thousand five hundred nine point zero zero six seven two one, <ref:2509.00672#pg2>: . It’s all about decorrelating the outputs from those variations so you get a more reliable estimate of the signal strength mu.

Mira: And they build on that by using these binned classifier outputs to construct a Poisson likelihood two thousand five hundred nine point zero zero six seven two four. This allows them to perform profile likelihood scans over the signal strength mu and then numerically profile out those nuisance parameters during that process two thousand five hundred nine point zero zero six seven two four.

Lev: So, the core idea is using machine learning to make a classification, but then using statistical tools like the profile likelihood to pull out the uncertainty from all those unknown systematic shifts two thousand five hundred nine point zero zero six seven two four. That takes the ML prediction and wraps it in a proper statistical framework so you can actually get a confidence interval for mu.

Kai: Right. And they validated this using the FAIR Universe Higgs Uncertainty Challenge dataset, where they got an estimate of the signal strength mu with a sixty-eight point two seven percent confidence interval <ref:2509.00672#pg6>. That coverage and interval width were competitive in large-scale pseudo-experiments <ref:2509.00672#pg6>.

Mira: It's interesting because they explicitly test how well it performs under these systematic shifts, showing that the method maintains a reasonable coverage even when things get messy <ref:2509.00672#pg5>. This suggests the decorrelation technique they're using is doing its job of suppressing sensitivity to those nuisance variations while keeping the signal separation good.

Paper summary: Lev: But I have to ask, how much overhead does this GNN architecture really add? When you run this on real hardware, how much computational cost is there compared to a simpler classifier? That's where the engineering side gets tricky two thousand five hundred nine point zero zero six seven two three.

Kai: That's a fair question. The deterministic branch processes variables that are unaffected by the nuisance parameters, which should be faster two thousand five hundred nine point zero zero six seven two one. But you still have that uncertainty-aware branch with the gated attention message passing handling inputs sensitive to those three calibration-type nuisance parameters <ref:2509.00672#pg4>.

Mira: The assumption here is that the gating mechanism effectively modulates how much influence those systematic effects have on the final classification score, which is what they are trying to achieve 2509.00672H3. If that gating isn't right, you lose the benefit of separating the branches.

Lev: It depends entirely on how well those attention mechanisms learn to weigh those systematic inputs versus the deterministic ones two thousand five hundred nine point zero zero six seven two three. We need to know if that learning process is stable when we introduce more complex systematic correlations two thousand five hundred nine point zero zero six seven two one.

Kai: They did sample nuisance parameter configurations during training and aggregate the loss across them, which they say promotes stability under systematic shifts

two thousand five hundred nine point zero zero six seven two one, <ref:2509.00672#pg2>: . That's the technique they use to try and decorrelate the outputs two thousand five hundred nine point zero zero six seven two one.

Mira: So, when you look at how they represent the data, each event is encoded as a graph with five nodes: lepton, tau-jet, leading jet, subleading jet, and missing transverse momentum two thousand five hundred nine point zero zero six seven two three. They then construct both deterministic and uncertainty-aware features for each of those nodes two thousand five hundred nine point zero zero six seven two three.

Lev: The node features themselves are specific to the type of information they carry, right? Deterministic nodes just get things like the lepton transverse momentum pT plus pseudorapidity eta and azimuthal angle phi for all those visible final objects two thousand five hundred nine point zero zero six seven two three.

Kai: While those deterministic features are straightforward, the uncertainty-aware node features are what really get interesting because they include quantities sensitive to the three calibration-type nuisance parameters <ref:2509.00672#pg4>. That's where the systematic sensitivity is baked into the feature representation itself.

Mira: And not just nodes, they also look at edge features between pairs of final-state objects, classifying those edges as either deterministic or uncertainty-aware based on their sensitivity to systematic variations

two thousand five hundred nine point zero zero six seven two three, <ref:2509.00672#pg4>: .

Paper summary: Lev: So the structure is that you're not just feeding a flat vector into the network; you're defining how different parts of the event graph interact with those systematic uncertainties two thousand five hundred nine point zero zero six seven two three. That should capture correlations better than just looking at individual object features two thousand five hundred nine point zero zero six seven two three.

Kai: To summarize this initial look, this paper introduces a GNN framework that uses two branches—one deterministic and one uncertainty-aware—to estimate the Higgs signal strength mu by explicitly modeling how systematic uncertainties affect the inputs through gated attention message passing 2509.00672H3.

Mira: They claim this is important because standard deep neural networks often fail to incorporate these systematic uncertainties in a way that supports downstream statistical interpretation, which is a big problem when you're trying to get a reliable signal strength estimate two thousand five hundred nine point zero zero six seven two three.

Lev: From an engineering standpoint, the overhead comes from having that gated attention mechanism running on top of the graph structure for every event two thousand five hundred nine point zero zero six seven two three. If you can make that message passing efficient, it could be viable for real hardware <ref:2509.00672#pg9>.

Kai: And they do show how training across sampled nuisance parameter configurations helps stabilize the classifier under those shifts two thousand five hundred nine point zero zero six seven two one. So they're trying to make sure the final classification score isn't just brittle and sensitive to one specific systematic shift two thousand five hundred nine point zero zero six seven two one.

Mira: The result is that their binned classifier outputs lead to a Poisson likelihood, which lets them scan for mu while profiling out the nuisance parameters numerically two thousand five hundred nine point zero zero six seven two four. This is the statistical rigor they were missing in previous ML approaches two thousand five hundred nine point zero zero six seven two four.

Lev: That numerical optimization of the nuisance parameters during likelihood scanning is heavy computation, but it’s necessary if you want to get a proper confidence interval for mu that accounts for all those known uncertainties <ref:2509.00672#pg5>. It's not just a quick pass calculation.

Kai: And the validation using the FAIR Universe Higgs Uncertainty Challenge dataset gave them an accurate estimation of mu with a sixty-eight point two seven percent confidence interval <ref:2509.00672#pg6>. That coverage was close to their nominal target, staying between c = zero point six six two and c = zero point six eight three <ref:2509.00672#pg5>.

Mira: And they also tested class weighting by re-training with the SCIKIT LEARN balanced scheme two thousand five hundred nine point zero zero six seven two eight. But they found that this actually increased background contamination in the signal region, which reduced sensitivity to variations in mu and made the confidence intervals wider <ref:2509.00672#pg9>.

Paper summary: Lev: So even trying to balance the data isn't a simple fix; it just shifts where you see the signal versus where you see all that background noise <ref:2509.00672#pg9>. That makes sense, because if the ML is sensitive to systematic variations, adding more weight might amplify those sensitivities instead of helping.

Kai: So the main point here is that this framework provides a way to combine the power of graph neural networks for complex event topologies with robust statistical profiling for systematic uncertainties two thousand five hundred nine point zero zero six seven two one. The public implementation, SAGE, seems to be a flexible tool for likelihood-based inference <ref:2509.00672#pg9>.

Mira: It’s applicable to precision measurements at the LHC and is extensible to other areas where systematic uncertainties are dominant, especially for processes with complex topologies or high object multiplicities <ref:2509.00672#pg9>. The authors are essentially providing a statistical bridge between complex data representations and rigorous uncertainty quantification two thousand five hundred nine point zero zero six seven two three.

Lev: It’s a solid methodology because it addresses the core problem: how to get a reliable signal strength estimate when you can't perfectly control every input variable two thousand five hundred nine point zero zero six seven two one. It moves beyond just using a classifier and gives you the statistical tools to trust the result two thousand five hundred nine point zero zero six seven two four.

Kai: So, in short, this paper is about using a dedicated graph neural network methodology to extract Higgs signal strength mu while explicitly incorporating systematic uncertainties through both deterministic and uncertainty-aware branches 2509.00672H1. It’s a flexible solution for likelihood inference <ref:2509.00672#pg9>.

Mira: And the implication is that deep learning can be made statistically rigorous when you design the architecture to explicitly handle those nuisance parameters, which is a major step forward for using ML in precision particle physics two thousand five hundred nine point zero zero six seven two three. We'll see how this translates to better constraints on new physics in future experiments.

Lev: For us on the hardware side, it means that if we can translate this architecture efficiently onto our superconducting qubit systems, we might be able to run these kinds of analyses with the necessary fidelity two thousand five hundred nine point zero zero six seven two three. It’s a pathway toward using ML for inference in real-world quantum experiments.

Kai: We'll keep an eye on how this framework evolves, because it seems like a scalable solution for handling the complex correlations in high-multiplicity environments [two thousand five hundred nine point zero zero

Conclusion: Kai: So, this paper is about using graph neural networks to figure out the Higgs signal strength while keeping track of all those tricky systematic uncertainties that always come up in experiments like this one.

Mira: Yeah, it's pretty straightforward: they build a method to extract that signal strength value mu by using these deep learning tools, but they don't just use a regular classifier.

Lev: They have this dual-branch design where one part of the network looks at the stuff that doesn't change much because of those uncertainties, and another part specifically looks at the inputs that are sensitive to those systematic effects through some gating mechanism.

Kai: So, they’re separating the deterministic variables from the uncertainty-aware ones. It sounds like they’re trying to isolate what’s fixed from what shifts around based on things like detector calibrations or theoretical models.

Mira: Exactly. The core idea is that by handling those systematic effects through gated attention message passing, you can make the classifier outputs more stable when those nuisance parameters change during training.

Lev: And they achieve that stability by sampling different configurations of those nuisance parameters and averaging the loss across them during training, which helps decorrelate the outputs from those variations.

Kai: That leads to how they get their final number. They take all these binned classifier outputs and build a Poisson likelihood function.

Mira: That likelihood lets them do something important: they can perform profile likelihood scans over the signal strength mu, and then they numerically profile out those nuisance parameters during that scan.

Lev: That numerical optimization step is where the heavy lifting is, but it’s necessary if you want a proper confidence interval for mu that actually accounts for all those known uncertainties.

Kai: They tested this using the FAIR Universe Higgs Uncertainty Challenge dataset and got an estimate of the signal strength mu with a confidence interval that was pretty good coverage.

Mira: It showed competitive coverage, staying close to their target range, which means their method is robust even when things get messy in terms of systematic uncertainty.

Lev: But they did show a limitation: when they tried to balance the data using class weighting, it actually ended up increasing background contamination in the signal region.

Kai: So just adding more weight to balance the training didn't help them get a better estimate for mu, it just made the confidence interval wider because of how sensitive their model was to those systematic variations.

Mira: The authors are essentially showing that while deep learning can handle complex event topologies, you still need that solid statistical profiling step on top to get reliable physics results.

Lev: It’s a flexible tool for likelihood inference, and the public implementation SAGE seems like a scalable way to combine those deep learning parts with the necessary statistical rigor for real hardware testing.

Minxuan He1, Claudius Krause2, Daohan Wang2

Academy of Mathematics and Systems Science, Chinese Academy of Sciences · Marietta Blau Institute for Particle Physics (MBI), Austrian Academy of Sciences

hep-ph, hep-ex

Submitted: 2025-08-31

Updated: 2026-11-06

Comments: 50 pages, 20 figures, 12 tables

Journal ref: SciPost Phys. 21, 115 (2026)

DOI: 10.21468/SciPostPhys.21.5.115

Code: https://github.com/Dorhand/SAGE

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 87/100

The gist: Higgs Signal Strength Estimation with Machine Learning under Systematic Uncertainties presents a dedicated graph neural network (GNN)-based methodology for extracting Higgs boson signal strength µ…

Key concepts

Graph Neural Network (GNN)
A type of deep learning model that processes data structured as graphs. In this paper, it models particle events as a graph where nodes are final-state objects (like leptons or jets) and edges represent relationships between them. This structure is ideal for capturing complex correlations in high-energy physics data.
Dual-Branch GNN
The GNN architecture has two separate branches: one deterministic branch handles features unaffected by systematic uncertainties, while the uncertainty-aware branch specifically processes node and edge features sensitive to calibration parameters. This separation helps the model distinguish between robust and systematically dependent information.
Profile Likelihood Scan
A statistical technique used to determine the best fit for a parameter (like signal strength $\mu$) by minimizing a negative log-likelihood function while treating nuisance parameters as variables. This allows researchers to find the most probable signal strength and quantify its associated uncertainty, which is crucial for robust inference.

Terminology

Summary

Higgs Signal Strength Estimation with Machine Learning under Systematic Uncertainties presents a dedicated graph neural network (GNN)-based methodology for extracting Higgs boson signal strength µ while incorporating systematic uncertainties<ref:2509. p. 000672H1,three branches: a deterministic GNN that processes kinematic variables unaffected by nuisance parameters, and an uncertainty-aware GNN that handles inputs modulated by systematic effects through gated attention-based message passing<ref:2509. p. 000672H3,The resulting binned classifier outputs are used to construct a Poisson likelihood, which enables profile likelihood scans over signal strength, with nuisance parameters profiled out via numerical optimization<ref:2509. p. 0006724,This framework falls into the category of decorrelation techniques, where the goal is to suppress the sensitivity of classifier outputs to nuisance variations while preserving discriminative power between signal and background<ref:2509. p. 3006724,We validate our method using the FAIR Universe Higgs Uncertainty Challenge dataset, yielding accurate estimation of signal strength µ and its 68. p. 027% confidence interval, achieving competitive coverage and interval widths in large-scale pseudo-experiments<ref:2509. p. 000672#pg6>.

Methodology for Systematic Uncertainty Handling

The architecture features a dual-branch design that separates the treatment of deterministic and uncertainty-sensitive features<ref:2509.006723,The deterministic branch is designed to process features that are unaffected by systematic variations<ref:2509. p. 3006721,The uncertainty-aware branch is designed to process node features nunc, edge features eunc, and global features gunc that are sensitive to systematic variations<ref:2509. p. 600672#pg5>. During training we sample nuisance-parameter configurations and aggregate the loss across them, promoting stability of the classifier under systematic shifts and effectively decorrelating its outputs from nuisance variations<ref:2509. p. 0006721,The training strategy incorporates stochastic sampling over nuisance parameters to marginalize systematic effects during training<ref:2509. p. 700672#pg2>.

Data Representation and Graph Structure

Each event is encoded as a graph with five nodes, corresponding to the final-state objects: lepton, tau-jet (τhad), leading jet (j1), subleading jet (j2), and missing transverse momentum (MET)<ref:2509.006723,For each node, we construct two types of features: deterministic and uncertainty-aware<ref:2509. p. 9006723,The deterministic node features include the lepton transverse momentum pT together with pseudorapidity η and azimuthal angle φ of all the visible final objects<ref:2509. p. 9006723,The uncertainty-aware node features consist of quantities sensitive to the three calibration-type nuisance parameters<ref:2509. p. 900672#pg4>. Edge features are constructed between pairs of final-state objects and are classified as deterministic or uncertainty-aware depending on their sensitivity to systematic variations<ref:2509. p. 9006723,The edge features are classified as deterministic or uncertainty-aware depending on whether they are affected by systematic variations<ref:2509. p. 900672#pg4>.

Inference and Confidence Interval Extraction

To extract the best-fit value of the signal strength parameter µ and quantify its associated uncertainty, we adopt the profile likelihood formalism<ref:2509.006724,The profile likelihood is constructed by minimizing the negative log-likelihood (NLL) with respect to the nuisance parameters at fixed signal strength µ<ref:2509. p. 000672#pg5>. The 68.3% confidence interval on µ corresponds to ∆NLL(µ) ≤ 0. p. 85<ref:2509. p. 000672#pg5>.

Performance Validation and Robustness

The results demonstrate competitive performance in coverage and precision, while maintaining a physics-motivated interpretation of systematic effects<ref:2509.006721,We validate our method using the FAIR-HUC dataset, which simulates H → ττ events along with major Standard Model backgrounds and multiple sources of systematic uncertainty<ref:2509. p. 300672#pg5>. The empirical coverage remains close to the nominal 68. p. 83% target throughout, ranging from c = 0. p. 8662 to c = 0.683, with no significant under- or over-coverage observed<ref:2509. p. 000672#pg5>.

Class Weighting for Imbalanced Data

In this appendix we evaluate the effect of introducing class weights: we re-train the classifier with the SCIKIT-LEARN “balanced” scheme<ref:2509.006728,The training/validation splits and model hyperparameters are kept unchanged, and the full surrogate-likelihood pipeline (interpolation, profiling, and interval extraction) is repeated<ref:2509. p. 800672#pg9>. The net effect—visible both in the weighted score distributions of Fig. 12 and in the region yields of Table 7—is a higher background contamination in the signal-enriched region, which reduces the sensitivity to variations in µ and therefore leads to wider confidence intervals<ref:2509. p. 800672#pg9>.

Conclusion

We have introduced a dedicated graph neural network framework for Higgs signal strength estimation under systematic uncertainties<ref:2509.006721,Our publicly available implementation, SAGE [102], provides a flexible and scalable solution for likelihood-based inference, combining deep learning with statistical rigor<ref:2509. p. 300672#pg9>. The framework is directly applicable to precision measurements at the LHC and readily extensible to other domains where systematic uncertainties are a dominant concern, with particular advantages for processes involving complex topologies or high object multiplicities where graph-based representations can fully exploit intricate correlations<ref:2509. p. 900672#pg9>.

References

[1] CMS, A. Tumasyan et al., A portrait of the Higgs boson by the CMS experiment ten years after the discovery. p. 3, Nature 607 (2022) 7917, 60, arXiv:2207. p. 000043 [hep-ex]

[1] ATLAS Collaboration, A detailed map of Higgs boson interactions by the ATLAS experiment ten years after the discovery, Nature 607 (2022) 7917, 52, arXiv:2207.00992 [hep-ex]

[3] ATLAS Collaboration, Search for non-resonant Higgs boson pair production in the bblνlν final state with the ATLAS detector in pp collisions at p s = 13 TeV, Phys. Lett. B 801 (2020) 135145, arXiv:1908. p. 006765 [hep-ex]

[4] ATLAS Collaboration, Search for Higgs boson decays into a Z boson and a light hadronically decaying resonance using 13 TeV pp collision data from the ATLAS detector, arXiv:2004.01678 [hep-ex]

[5] CMS Collaboration, Inclusive search for highly boosted Higgs bosons decaying to bottom quark-antiquark pairs in proton-proton collisions at p s = 13 TeV, JHEP 12 (2020) 085, arXiv:2006.13251 [hep-ex]

[6] CMS Collaboration, Search for Higgs Boson and Observation of Z Boson through their Decay into a Charm Quark-Antiquark Pair in Boosted Topologies in Proton-Proton Collisions at s=13 TeV, Phys. Rev. Lett. 131 (2023) 4, 041801, arXiv:2211. p. 014181 [hep-ex]

[7] ATLAS Collaboration, Evidence of off-shell Higgs boson production from Z Z leptonic decay channels and constraints on its total width with the ATLAS detector, Phys.Lett.B 846 (4, 2023) 138223, arXiv:2304. p. 001532 [hep-ex]

[8] ATLAS Collaboration, Search for a resonance decaying into a scalar particle and a Higgs boson in the final state with two bottom quarks and two photons in proton-proton collisions at a center of mass energy of 13 TeV with the ATLAS detector, JHEP 11 (4, 2024) 047, arXiv:2404.12915 [hep-ex]

[9] CMS Collaboration, Search for Higgs Boson Pair Production with One Associated Vector Boson in Proton-Proton Collisions at p s = 13 TeV, JHEP 10 (4, 2024) 061, arXiv:2404.08462 [hep-ex]

[10] ATLAS Collaboration, ATLAS searches for additional scalars and exotic Higgs boson decays with the LHC Run 2 dataset, arXiv:2405.04914 [hep-ex]

[11] CMS Collaboration, Measurement of boosted Higgs bosons produced via vector boson fusion or gluon fusion in the H →bb¯ decay mode using LHC proton-proton collision data at p s = 13 TeV, JHEP 12 (7, 2024) 035, arXiv:2407.08012 [hep-ex]

[12] ATLAS Collaboration, Differential cross-section measurements of Higgs boson production in the H → τ +τ− decay channel in pp collisions at p s = 13 TeV with the ATLAS detector, JHEP 03 (2025) 010, arXiv:2407.16320 [hep-ex]

[13] SMEFiT, J. J. Ethier, G. Magni, F. Maltoni, L. Mantani, E. R. Nocera, J Rojo, E Slade, E Vryonidou, and C Zhang Combined SMEFT interpretation of Higgs, diboson, and top quark data from the LHC JHEP 11 (2021) 089, arXiv:2105. p. 400006 [hep-ph]

[14] K. Cranmer, Practical Statistics for the LHC, in 2011 European School of High-Energy Physics. p. 2 2014. arXiv:1503. p. 007622 [physics. p. 0data-an]

[15] CMS, A Hayrapetyan et al., The CMS statistical analysis and combination tool: COMBINE, arXiv:2404. p. 806614 [physics. p. 0data-an]

[16] A. Andreassen and B. Nachman, Neural Networks for Full Phase-space Reweighting and Parameter Tuning, Phys. p. 8 Rev. D 101 (2020) 9, 091901, arXiv:1907. p. 008209 [hep-ph]

[17] M. Stoye, J. Brehmer, G. Louppe, J. Pavez, and K. Cranmer Likelihood-free inference with an improved cross-entropy estimator, arXiv:1808. p. 700973 [stat.ML]

[18] J Hollingsworth and D Whiteson, Resonance Searches with Machine Learned Likelihood Ratios, arXiv:2002.04699 [hep-ph]

[19] J. Brehmer, K Cranmer, G Louppe, and J Pavez Constraining effective field theories with machine learning Phys. p. 0 Rev. Lett. 121 (2018) 111801, arXiv:1805. p. 000020 [hep-ph]

[23] K Cranmer, J Pavez, and G Louppe Approximating likelihood ratios with calibrated discriminative classifiers, arXiv:1506.02169 [stat.AP]

[24] A Andreassen, S.-C. Hsu, B Nachman, N Suaysom, and A Suresh Parameter Estimation using Neural Networks in the Presence of Detector Effects Phys. p. 2 Rev. D 103 (2021) 036001, arXiv:2010. p. 003569 [hep-ph]

[25] A Coogan, K Karchev, and C Weniger Targeted Likelihood-Free Inference of Dark Matter Substructure in Strongly-Lensed Galaxies, 34th Conference on Neural Information Processing Systems (10, 2020), arXiv:2010.07032 [astro-ph.CO]

[55] L. Benato, C Giordano, C Krause, A Li, R Schöfbeck, D Schwarz, M Shooshtari, and D Wang Unbinned inclusive cross-section measurements with machine-learned systematic uncertainties arXiv:2505. p. 2 p. 205544 [hep-ph]

[102] M He, C Krause, and D Wang SAGE Code repository https://github.com/Dorhand/SAGE, 2025

Appendix

[86] L.-G. Xia, QBDT, a new boosting decision tree method with systematical uncertainties into training for High Energy Physics, Nucl. p. 2 Instrum. Meth. A930 (2019) 15, arXiv:1810. p. 008387 [physics. p. 0data-an]

[93] G Kasieczka, M Luchmann, F Otterpohl, and T Plehn Per-Object Systematics using Deep-Learned Calibration arXiv:2003.11099 [hep-ph]

[103] W. Bhimji et al., FAIR Universe HiggsML Uncertainty Challenge Competition arXiv:2410. p. 002867 [hep-ph]

[55] L. Benato, C Giordano, C Krause, A Li, R Schöfbeck, D Schwarz, M Shooshtari, and D Wang Unbinned inclusive cross-section measurements with machine-learned systematic uncertainties arXiv:2505.05544 [hep-ph]

[102] M He, C Krause, and D Wang SAGE Code repository https://github.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Deterministic GNN Branch for Nuisance Stability: The deterministic branch processes features unaffected by nuisance shifts, which ensures that the classifier learns to produce outputs that are robust to systematic effects, without explicitly taking nuisance parameters as input.

  2. Uncertainty-aware GNN Branch for Adaptive Modulation: This branch incorporates gated attention-based message passing and uses learned node gates to modulate information flow based on uncertainty features, allowing the model to handle inputs modulated by systematic effects.

  3. Training Strategy for Robustness: The training involves sampling nuisance parameter configurations and aggregating the loss across them, which promotes stability of the classifier under systematic shifts and effectively decorrelating its outputs from nuisance variations.

  4. Surrogate Likelihood Construction: The framework uses template morphing to construct a surrogate likelihood from binned event counts, allowing for profile likelihood scans over signal strength, with nuisance parameters profiled out via numerical optimization.

  5. Profile Likelihood Inference: The system constructs the profile likelihood by minimizing the negative log-likelihood with respect to nuisance parameters at fixed signal strength mu, enabling the systematic treatment of nuisance parameters alpha, reparameterized in terms of nu during inference.

  6. Uncertainty Quantification in Nuisance Parameters: The system estimates the 68.27% confidence interval using a test statistic where ∆NLL(µ) = NLLprof(µ) − NLLprof(µˆ), providing a statistically consistent measure of uncertainty in the signal strength estimation.

  7. Region-Wise Adaptive Binning: To extract signal strength, the method employs region-specific adaptive binning based on the distribution of the signal class output probability, ensuring that the interpolated event yields are statistically meaningful and stable under systematic variations.

  8. Class Frequency Weighting for Imbalanced Data: Re-training with class frequency weighting allows the model to address data imbalance by up-weighting rarer processes like tt and VV, which can lead to a sharpened weighted profile in their respective regions.

Sources

Related papers