Potential and limitations of machine-learning approaches to inclusive $|V_{ub}|$ determinations

Abstract

The determination of $|V_{ub}|$ in inclusive semileptonic $B \to X_u \ell \nu$ decays will be among the pivotal tasks of Belle II. In this paper we study the potential and limitations of machine learning approaches that attempt to reduce theory uncertainties by extending the experimentally accessible fiducial region of the $B\to X_u \ell\nu$ signal into regions where the $B\to X_c \ell\nu$ background is dominant. We find that a deep neural network trained on low-level single particle features offers modest improvement in separating signal from background, compared to BDT set-ups using physicist-engineered high-level features. We further illustrate that while the signal acceptance of such a deep neural network deteriorates in kinematic regions where the signal is small, such as at high hadronic invariant mass, neural networks which exclude kinematic features are flatter in kinematics but less inclusive in the sampling of exclusive hadronic final states at fixed kinematics. The trade-off between these two set-ups is somewhat Monte Carlo dependent, and we study this issue using the multipurpose event generator Sherpa in addition to the widely used $B$-physics tool EvtGen.