Abstract
We implement a convolutional neural network (CNN) to study the $\Sigma$ hyperons using experimental data of the $K^-p \to \pi^0 \Lambda$ reaction. The averaged accuracy of the NN models in resolving resonances on the test data sets is $98.5\%$, $94.8\%$, and $82.5\%$ for one-, two-, and three-additional-resonance cases, respectively. We find that the three most significant resonances are $1/2^+$, $3/2^+$, and $3/2^-$ states with masses of $1.62(11)$ GeV, $1.72(6)$ GeV, and $1.61(9)$ GeV, and probabilities of $100(3)\%$, $72(24)\%$, and $98(52)\%$, respectively. Our results support the three-star $\Sigma(1660)1/2^+$, the one-star $\Sigma(1780)3/2^+$, and the one-star $\Sigma(1580)3/2^-$ in the Particle Data Group (PDG) rankings. The ability to provide quantitative probabilities and numerical stability makes NN a potentially transformative tool in baryon partial wave analysis.
Introduction
In hadron spectroscopy, researchers study how quarks bind to form particles like protons and hyperons ($\Sigma, \Lambda, \Xi$). Identifying these particles typically involves Partial Wave Analysis (PWA)—fitting theoretical models to experimental scattering data. However, traditional PWA based on $\chi^2$ minimization faces two major hurdles: it lacks a clear way to assign existence probabilities to resonances, and fits often become unstable when accounting for multiple overlapping states. Consequently, many predicted $\Sigma$ resonances remain "poorly established" (1-star or 2-star) in the PDG. This work utilizes deep learning to provide clarity and statistical confidence to these elusive states.
Methodology
The researchers implemented a CNN architecture for a dual task: Classification (CA) to identify quantum numbers ($J^P$) and Regression (RE) to determine physical properties (mass and width).
To overcome the scarcity of experimental data, the team used "Effective Lagrangian" theory to generate a massive synthetic dataset of 20 million scattering events for training. They employed Artificially Expanded Data (AED), distorting training sets with noise to mimic real-world experimental uncertainties. The analysis proceeded sequentially, examining cases with one, two, and three additional resonances to ensure model stability as complexity increased.
Key Results and Discussion
The NN models achieved high accuracy on test datasets: $98.5\%$ for the 1-resonance case and $82.5\%$ for the 3-resonance case. Unlike traditional fits that provide only a "best value," the NN yields a quantitative probability for each state, which is crucial for classifying star-rated particles.
The analysis identified three dominant resonances that align well with existing PDG data:
- $1/2^+$ state: Mass of $1.62(11)$ GeV with $100(3)\%$ probability (matching the 3-star $\Sigma(1660)$).
- $3/2^+$ state: Mass of $1.72(6)$ GeV with $72(24)\%$ probability (supporting the 1-star $\Sigma(1780)$).
- $3/2^-$ state: Mass of $1.61(9)$ GeV with $98(52)\%$ probability (supporting the 1-star $\Sigma(1580)$).
Conversely, the probability for a $1/2^-$ state was found to be consistent with zero, suggesting it does not play a significant role in the studied energy range. The NN proved remarkably stable; parameters for established resonances did not fluctuate wildly when additional states were added to the model.
Conclusion and Future Outlook
This study demonstrates that NN-based PWA can successfully resolve subatomic resonances with high statistical rigor. While currently limited to a single reaction channel ($K^-p \to \pi^0 \Lambda$), the approach is easily extendable to multi-channel analyses. Future work will apply this method to upcoming high-precision data from facilities like J-PARC, JLAB, and the future PANDA experiment to further map the baryon spectrum.
Summary updated