Unfolding Particle Physics Data with Machine Learning: Diverse Techniques and Benchmarks

Core Concepts

Machine learning techniques, including reweighting, distribution mapping, and generative unfolding, can accurately unfold particle physics data from detector-level to particle-level, enabling detailed measurements of the Standard Model and sensitivity to new phenomena.

Abstract

The paper introduces and benchmarks a diverse set of machine learning-based unfolding methods on two particle physics datasets: Z+jets production and top quark pair production.
Key highlights:

Reweighting methods like OmniFold and its Bayesian variant bOmniFold can accurately reproduce particle-level spectra across complex observables.
Distribution mapping approaches like Schrödinger Bridge and Direct Diffusion also demonstrate strong unfolding performance.
Generative unfolding networks, including cINN, Transfermer, CFM, TraCFM, and Latent Diffusion, can learn the conditional probability to unfold from detector-level to particle-level.
All techniques offer an exciting toolkit for a new class of particle physics measurements with unprecedented detail and sensitivity to new phenomena.
The authors benchmark the methods on the same datasets, facilitating direct comparisons and highlighting the diverse strengths of the approaches.

Stats

The Z+jets dataset contains around 24M simulated events, with 20M for training and 4M for testing.
The top pair dataset is not specified in terms of event counts.

Quotes

"Recent innovations from machine learning allow for data unfolding, without binning and including correlations across many dimensions."
"All techniques are capable of accurately reproducing the particle-level spectra across complex observables."
"Given that these approaches are conceptually diverse, they offer an exciting toolkit for a new class of measurements that can probe the Standard Model with an unprecedented level of detail and may enable sensitivity to new phenomena."

Key Insights Distilled From

The Landscape of Unfolding with Machine Learning

by Nath... at arxiv.org 04-30-2024

https://arxiv.org/pdf/2404.18807.pdf

Deeper Inquiries

How can the model dependence of the unfolding be further reduced, beyond the single-step approaches presented here

To further reduce the model dependence of the unfolding process, several strategies can be implemented beyond the single-step approaches discussed in the context.

Iterative Unfolding: Instead of performing unfolding in a single step, iterative unfolding can be employed. This involves iteratively updating the unfolded distribution based on the previous iteration's result. By incorporating feedback from the unfolding process itself, the model dependence can be reduced as the algorithm refines its estimate with each iteration.

Regularization Techniques: Including regularization techniques in the unfolding process can help mitigate model dependence. Regularization methods such as dropout, weight decay, or early stopping can prevent overfitting and improve the generalization of the unfolding algorithm to unseen data.

Ensemble Methods: Utilizing ensemble methods by combining the results of multiple unfolding algorithms can help reduce model dependence. By averaging the outputs of different unfolding techniques, the final unfolded distribution can be more robust and less sensitive to the specific assumptions of any single method.

Incorporating Prior Knowledge: Incorporating prior knowledge about the underlying physics processes can also help reduce model dependence. By constraining the unfolding process with known physical principles or relationships, the algorithm can be guided towards more accurate and reliable results.

What are the implications of the different unfolding methods for combining data from multiple experiments in global analyses

The implications of different unfolding methods for combining data from multiple experiments in global analyses are significant.

Consistency in Unfolding: Different unfolding methods may introduce biases or uncertainties in the unfolded distributions. When combining data from multiple experiments, it is crucial to ensure that the unfolding techniques used are consistent across all datasets to avoid introducing inconsistencies or artifacts in the combined analysis.

Cross-Validation: Cross-validating the unfolding results from different methods on overlapping datasets can help assess the compatibility and consistency of the unfolded distributions. This validation process is essential when combining data from various experiments to ensure the reliability of the combined analysis.

Method Selection: Choosing the appropriate unfolding method for each dataset based on its characteristics and uncertainties is crucial for global analyses. Understanding the strengths and limitations of each unfolding technique and selecting the most suitable method for each dataset can improve the overall quality and reliability of the combined analysis.

Uncertainty Quantification: Different unfolding methods may yield varying levels of uncertainty in the unfolded distributions. When combining data, it is essential to properly quantify and propagate these uncertainties to provide accurate and reliable results in global analyses.

How can the unfolding techniques be extended to handle systematic uncertainties in the detector simulation and the underlying theory predictions

Extending unfolding techniques to handle systematic uncertainties in the detector simulation and underlying theory predictions is essential for robust data analysis in particle physics.

Systematic Uncertainty Modeling: Including systematic uncertainties in the unfolding process requires incorporating variations in the detector response and theoretical predictions. This can be achieved by introducing nuisance parameters that account for uncertainties in the simulation models and theory predictions.

Uncertainty Propagation: Propagating systematic uncertainties through the unfolding process involves considering how variations in the input parameters affect the unfolded distributions. Techniques such as Monte Carlo sampling or Bayesian methods can be used to quantify and propagate these uncertainties through the unfolding algorithm.

Systematic Error Estimation: Properly estimating systematic errors in the detector simulation and theory predictions is crucial for accurate unfolding. This involves studying the impact of systematic variations on the unfolded results and providing reliable estimates of the associated uncertainties.

Robust Validation: Validating the unfolding techniques under different systematic scenarios is essential to ensure the reliability and robustness of the results. Performing sensitivity studies and robustness checks can help assess the impact of systematic uncertainties on the unfolded distributions and improve the overall quality of the analysis.