Abstract
A deep learning method using a Convolutional Neural Network (CNN) is developed to determine impact parameters in heavy-ion collisions at low-to-intermediate energies (50–100 MeV/nucleon). By treating fragment momentum distributions as images, the CNN overcomes the limitations of traditional scalar proxies. Improvements in input selection and training, including the consideration of experimental filter effects, allow the CNN to significantly outperform conventional methods, particularly in identifying central collision events and accurately deducing nuclear stopping power.
Introduction
In heavy-ion physics, the impact parameter ($b$)—the perpendicular distance between the centers of two colliding nuclei—is a fundamental variable that governs reaction dynamics, nuclear matter density, and particle production. However, $b$ cannot be measured directly. Experimenters must rely on observable proxies like the number of charged particles ($N_{ch}$) or the charge of the largest fragment ($Z_{max}$). At low-to-intermediate energies (50–100 MeV/nucleon), these traditional proxies often become "blunt instruments," struggling to distinguish between central and mid-range collisions. This paper presents a CNN-based approach that treats fragment momentum distributions as high-dimensional "images" to predict $b$ with high precision.
Methodology: Momentum Distributions as Images
The researchers moved beyond scalar proxies by utilizing the entire "shape" of the collision's output. Momentum data for all fragments were converted into 2D histograms (images) plotting absolute transverse momentum ($|P_t|/A$) against longitudinal momentum ($|P_z|/A$).
As shown in Figure 1, these distributions act as fingerprints. Separate channels for light fragments ($Z \le 2$) and heavy fragments ($Z > 2$) capture nuanced physics that single-number observables miss. The CNN acts as a pattern recognition tool, identifying subtle correlations in how particles are sprayed into the detector.
CNN Architecture and Training
The model was trained on 400,000 simulated $^{124}\text{Sn} + ^{124}\text{Sn}$ collision events generated via the Constrained Molecular Dynamics (CoMD) model.
The CNN architecture (Figure 2) features two convolutional layers with 32 kernels each, max-pooling for dimensionality reduction, and dropout layers to prevent overfitting. Crucially, the training process incorporated an "experimental filter" based on the INDRA detector array. This accounts for detector blind spots and energy thresholds, ensuring the model remains accurate when applied to real-world experimental data.
Key Results
- Superior Accuracy: The CNN achieved a mean absolute error ($\Delta b$) of 0.38 to 0.44 fm, significantly outperforming traditional methods which often exhibit errors twice as large in the central collision region.
- Central Collision Recognition: The CNN's ability to identify central events ($b/b_{max} < 0.2$) reached approximately 66% at 100 MeV/nucleon, a major improvement over conventional proxies that typically saturate at low $b$.
- Impact on Stopping Power: When used to select events for calculating nuclear stopping power ($R_E^p$), the CNN-derived values were much closer to theoretical "true" values than those from traditional selection methods.
Conclusion
The application of deep learning to impact parameter determination represents a significant advancement in experimental nuclear physics. By leveraging the full information content of momentum distributions, the CNN method provides higher resolution for central collisions and more reliable deductions of nuclear observables like stopping power. Future work will focus on cross-training with different transport models to ensure the network's predictions are robust and model-independent.
Summary updated