Exploring the use of machine learning in unfolding experimental data

Experimental measurements of steeply falling spectrum, such as jet momentum, must be corrected for the detector response, which shifts and smears the spectrum. Although the detector response is well understood, it is nontrivial to correct measured spectrum for the response. Traditional methods, such as Iterative Bayesian Unfolding (IBU), apply the detector response to simulated data and use the resulting spectrum to unfold the data bin by bin, approximating the true spectrum. Machine learning presents a new method of unfolding that can unfold data event by event, and is well suited to unfold multiple variables simultaneously. We investigate new unfolding methods for momentum spectrum, utilizing machine learning. In particular, we compare the results of IBU, implemented through RooUnfold, to the Omnifold algorithm. Tests are performed using Monte Carlo data. Finally, we implement the Omnifold algorithm using neural networks created in both Keras and XGBoost and compare the results to traditional methods.