Hybrid tree-based machine learning for calorimetric energy calibration using the Lorenzetti Showers framework

Résumé fourni par la source

Accurate and low-latency energy calibration is essential in high-energy physics calorimetry, especially in trigger-oriented environments with stringent timing constraints. Fast calorimetric measurements are affected by pile-up, electronic noise, detector nonlinearities, and shower fluctuations, degrading transverse-energy reconstruction and broadening threshold-based event selection. This work investigates hybrid tree-based machine-learning strategies for event-level calibration within a controlled Lorenzetti simulation benchmark. Datasets were generated with the Lorenzetti Showers framework configured as an ATLAS-inspired electromagnetic calorimeter, and ring-based shower observables were used as compact topological inputs. The proposed approach combines full-ring and dimensionality-reduced representations with Random Forest, Gradient Boosted Decision Tree, residual-learning, and stacked-ensemble regressors to estimate multiplicative calibration factors applied to the standard energy measurement. The best performance was achieved by a hybrid Random Forest model with Gradient Boosted Decision Tree residual correction using the full 100-ring representation. Under the considered benchmark conditions, this configuration reduced RMSE by 60.61%, IQR by 43.39%, relative energy resolution by 44.96%, and the threshold offset required to reach 90% event acceptance by 66.15%. The best compact configuration, a Gradient Boosted Decision Tree using five independent components, achieved reductions of 51.25% in RMSE, 25.45% in IQR, 31.98% in relative energy resolution, and 55.90% in the 90% acceptance threshold offset. The top-ranked configurations also showed low CPU-based trained-model prediction times. Overall, the results support ring-based hybrid tree calibration as a promising benchmark-level strategy for improving simulated fast calorimetric energy estimates, while validation with higher-fidelity simulations, experimental data, and target trigger hardware is required before deployment conclusions.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.