Machine-Learning-Inspired SMEFT Simplified Template Cross Sections: A Case Study in ZH Production

Abstract

The Simplified Template Cross Section (STXS) program has become the standard interface between Higgs measurements and global fits, but its fixed one-dimensional boundaries are not guaranteed to align with the phase-space directions to which the Standard Model Effective Field Theory (SMEFT) is most sensitive. We propose a machine-learning-inspired extension of STXS in which supervised classifiers are used only at the design stage to identify simple, publishable phase-space boundaries. Using associated Higgs production, $pp \to ZH$, as a case study and a benchmark momentum-dependent bosonic SMEFT deformation, we show that the relevant signal-background separation is well captured by a linear boundary in the $(p_T^Z,mZH)$ plane. We construct such boundaries with a linear support vector machine and with a deep-neural-network-assisted distillation procedure, and compare them directly with the standard STXS $p_T^Z$ bins through a common single-region Asimov-significance analysis. In this proof-of-concept setup, the ML-inspired regions systematically outperform the corresponding STXS regions, with the largest gains appearing in the boosted regime where SMEFT effects are concentrated. The final observable remains a simple linear cut, preserving the transparency and experimental portability that make STXS useful.

Figures (9)

• Figure 1: Asimov significance in the slope--intercept plane for the scan around the SVM solution in the highest-$p_T^Z$ slice, for a representative choice $\epsilon=0.5$.
• Figure 2: Classifier performance inside the four official STXS $p_T^Z$ slices. Left: ROC curves for SM--benchmark discrimination with a shallow classifier trained separately in each slice. Right: background efficiency versus signal efficiency. The increasing separability toward large $p_T^Z$ motivates the focus on the boosted region.
• Figure 3: Asimov significance as a function of the DNN-score threshold used to define the projected high-score region in the highest-$p_T^Z$ slice.
• Figure 4: Significance landscape for the post-fit slope--intercept scan applied to the DNN-distilled boundary in the highest-$p_T^Z$ slice, for $\epsilon=0.5$.
• Figure 5: Comparison of the official STXS boundary with the ML-inspired linear boundaries in the highest-$p_T^Z$ slice, shown for the representative background-uncertainty choices used in the analysis. The ML-guided lines select the correlated high-$p_T^Z$, high-$m_{ZH}$ region more efficiently than the one-dimensional STXS slicing.