Modern experimental physics research, particularly in particle physics, requires extensive data analysis efforts to identify significant signals indicative of new physics. We present BumpNet, a novel Neural Network (NN) architecture designed to conduct model-independent searches for mass bumps arising from new physics phenomena. This model maps invariant mass histograms into statistical inference distributions to facilitate efficient signal detection. By focusing on experimental data without relying on simulations, BumpNet enables the identification of exclusive selections that significantly deviate from the Standard Model’s known properties, marking them for further study. The NN minimizes resource-intensive tasks such as background estimation and systematic uncertainty evaluation, enabling rapid testing of multiple final states with only minor sensitivity loss compared to standard likelihood-based methods.

The model’s performance is validated using training data from the Dark Machines dataset, with its predicted significance benchmarked against an ideal likelihood analysis. The results demonstrate negligible bias and variance below 1$\sigma$ when tested on Gaussian-shaped signals. Furthermore, BumpNet’s consistency is evaluated using data from the ATLAS Higgs discovery, reinforcing its reliability and applicability in real-world analyses.