Physics-informed neural network (PINN) modeling of charged particle multiplicity using the two-component framework in heavy-ion collisions: A comparison with data-driven neural networks

This study demonstrates that a physics-informed neural network (PINN), which incorporates the two-component framework and hard-scattering constraints, outperforms conventional data-driven neural networks in predicting charged particle multiplicity in heavy-ion collisions, particularly in sparse high-multiplicity regions and when generalizing to unseen collision systems like Au+Au.

The Gist

Imagine you are trying to teach a computer to predict how many passengers will be on a crowded bus after a massive traffic jam. In the world of physics, this "bus" is a heavy-ion collision (like smashing two gold or zirconium nuclei together), and the "passengers" are the charged particles (hadrons) that fly out.

Physicists have a classic, rule-based way to guess this number, called the Glauber two-component formula. Think of this formula as a trusted, old-school recipe that says: "The total number of passengers is a mix of people who just bumped into each other gently (soft collisions) and people who crashed hard (hard collisions)."

However, in recent years, scientists have started using Neural Networks (NNs)—a type of artificial intelligence that learns by looking at millions of examples, much like a child learning to recognize cats by seeing thousands of pictures.

This paper compares two ways of teaching the AI to predict the number of particles:

1. The "Pure Data" Student (The Normal NN)

This is a standard AI. It is given a massive dataset of simulated collisions (specifically, 1 million collisions of Zirconium nuclei). It looks at the patterns, memorizes the relationship between the collision geometry and the number of particles, and tries to guess the answer for new situations.

• The Problem: It only knows what it has seen. If you ask it about a collision type it has never encountered (like Gold nuclei, which are bigger and produce more particles), it starts to guess wildly because it has no "common sense" or rules to fall back on. It's like a student who memorized the answers to a math test but doesn't understand the actual math, so they fail when the teacher changes the numbers.

2. The "Physics-Informed" Student (The PINN)

This is the star of the paper. The researchers didn't just let the AI look at data; they forced it to learn the old-school recipe (the Glauber formula) at the same time.

• How it works: Imagine the AI is taking a test. It gets points for getting the right answer based on the data, but it loses points if its answer violates the known physics rules. The AI has to find a balance: it must fit the data and obey the laws of physics.
• The Result: This AI actually learned a specific "secret ingredient" in the recipe (called $x$, the weight of hard collisions). It figured out that about 41% of the particles come from hard collisions. Because it understands the underlying rules, it doesn't just memorize; it understands the logic.

The Big Test: The "Unseen" Collision

The researchers put both AIs to the test with two new scenarios:

1. Ruthenium (Ru) collisions: These are "cousins" to the Zirconium they trained on (same size, just different chemistry).
   • Result: Both AIs did well. The "Pure Data" student could handle this because it was similar to what it studied.
2. Gold (Au) collisions: These are much bigger and produce way more particles than anything the AI saw during training. This is the "unseen" territory.
   • Result: The "Pure Data" student failed. It started underestimating the number of particles because it had never seen such high numbers before.
   • The Winner: The PINN (Physics-Informed) student did much better. Even though it had never seen Gold collisions, its knowledge of the physics rules allowed it to extrapolate (make a smart guess) into the unknown. It knew that if the collision is bigger, the number of particles must go up according to the rules, so it didn't get stuck.

Why This Matters

The paper shows that when you have limited data (or data that is sparse in certain areas, like very high-energy collisions), teaching the AI the rules of the game helps it learn faster and generalize better.

• The Analogy: If you teach a child to drive only by showing them videos of driving in the rain, they might panic when it's sunny. But if you teach them the rules of the road (stop at red lights, yield to pedestrians) alongside the videos, they can handle sunny days, snowy days, or even driving in a new city they've never visited.

Summary of Claims

• The researchers used a simulation model called HYDJET++ to generate 1 million training events.
• They successfully trained a PINN to extract the physical parameter $x$ (found to be ~0.41) directly from the data.
• The PINN outperformed the standard "Pure Data" AI, especially when predicting results for Gold (Au) collisions, which were completely new to the model.
• The study concludes that adding physical constraints acts as a "regularizer," helping the AI make better predictions even when training data is scarce or when facing new, unseen collision systems.

The paper does not claim to have solved all heavy-ion physics problems or to be ready for immediate clinical use; it is a proof-of-concept showing that mixing physics rules with AI makes the AI smarter and more reliable.

Technical Summary

Technical Summary: Physics-Informed Neural Network Modeling of Charged Particle Multiplicity in Heavy-Ion Collisions

Problem Statement
The study addresses the challenge of modeling charged hadron multiplicity ($N_{ch}$) in relativistic heavy-ion collisions, a key observable for understanding the initial geometry and particle production mechanisms of the quark-gluon plasma (QGP). While standard deep neural networks (NNs) have emerged as powerful tools for modeling complex nonlinear dependencies in high-energy physics, they rely solely on statistical correlations within training data. This data-driven approach lacks physical interoperability and often struggles with generalization, particularly in "sparse-data regimes" (e.g., high-multiplicity events) or when extrapolating to collision systems not present in the training set. The authors aim to determine whether integrating minimal, well-established physical constraints—specifically the Glauber two-component formula—into the neural network training process can improve model stability and generalization compared to purely data-driven approaches.

Methodology
The authors employed a comparative framework using two distinct neural network architectures trained on simulated data:

1. Data Generation: A dataset of one million minimum-bias $^{96}_{40}\text{Zr} + ^{96}_{40}\text{Zr}$ collision events was generated using the HYDJET++ event generator at $\sqrt{s_{NN}} = 200$ GeV. This model combines PYQUEN (for hard interactions and energy loss) and FASTMC (for thermal production), initialized via Glauber model calculations for participant nucleons ($N_{part}$) and binary collisions ($N_{coll}$).
2. Input/Output: The input vector $X = (b, N_{part}, N_{coll})$ (impact parameter, participants, binary collisions) was used to predict the target variable $Y = N_{ch}$.
3. Architectures:
   • Conventional NN (Normal NN): A standard feedforward network (3 input neurons, three hidden layers with 128, 64, and 32 neurons) trained solely to minimize the Mean Squared Error (MSE) between predicted and simulated $N_{ch}$.
   • Physics-Informed Neural Network (PINN): A parallel architecture incorporating a fixed analytical module based on the Glauber two-component formula:
     $$\frac{dN_{ch}}{d\eta} = n_{pp} \left[ (1 - x) \cdot \frac{N_{part}}{2} + x \cdot N_{coll} \right]$$
     Here, $n_{pp}$ (average charged hadrons in $pp$ collisions) was fixed at 3.602, while the hard-scattering fraction $x$ was treated as a trainable scalar parameter.
4. Loss Function: The PINN was trained using a composite loss function:
   $$\mathcal{L}_{total} = \mathcal{L}_{data} + \lambda \mathcal{L}_{physics}$$
   where $\mathcal{L}_{data}$ measures deviation from simulated event data, and $\mathcal{L}_{physics}$ penalizes deviations from the Glauber formula. The hyperparameter $\lambda$ controls the weight of the physical constraint.
5. Evaluation Strategy: Models were trained on Zr+Zr data and tested on:
   • Ru+Ru: An isobar system (seen test data).
   • Au+Au: A completely unseen system (unseen test data) with significantly higher multiplicities.
   • Sparse Data Regime: Specific attention was paid to high-$N_{ch}$ regions where training data is naturally scarce due to Glauber geometry statistics.

Key Results

• Parameter Extraction: The PINN successfully learned the hard-scattering fraction $x$ directly from the event data, converging to a value of $x \approx 0.41$. This value was validated by scanning fixed $x$ values, confirming that $0.41$ provided the best agreement with the simulated data.
• Impact of Physics Constraints ($\lambda$): Increasing $\lambda$ forced the PINN output to align more closely with the Glauber two-component formula. In regions of sparse data (high $N_{ch}$), where the data-driven loss provides weak gradient guidance, the physics constraint ensured physically consistent predictions.
• Generalization to Unseen Systems:
  • Ru+Ru (Isobar): Both the Normal NN and PINN achieved high accuracy, as Ru is an isobar of Zr with similar properties.
  • Au+Au (Unseen): The Normal NN significantly underpredicted multiplicities for Au+Au collisions (starting around $N_{ch} \approx 500$), failing to extrapolate beyond the training domain. In contrast, the PINN demonstrated superior generalization. With $\lambda = 1.0$, the PINN predicted higher $N_{ch}$ values (up to $\approx 1100$) that were absent in the Zr training set, though it did not fully reproduce the most central Au+Au distribution (which reaches $\sim 1750$).
• Limited Data Performance: When trained on a small subset of 500 data points, the PINN consistently outperformed the Normal NN across all metrics ($R^2$, RMSE, MAE), demonstrating that physics constraints act as an effective regularizer when training samples are scarce.

Significance and Claims
The paper claims that incorporating minimal, well-established physical constraints (the Glauber two-component formula) into neural network training significantly enhances the model's ability to generalize to unseen collision systems and sparse-data regimes. The authors emphasize that this study serves as an initial step to demonstrate the advantages of PINN methodology over purely data-driven networks in heavy-ion collision phenomenology.

The authors modestly state that the work is not intended to provide a final predictive description of all heavy-ion observables, but rather to highlight the role of PINNs in improving stability and extrapolation capabilities. They suggest that PINNs could be effective in reducing the number of simulated events required for analysis and decreasing overall computation time. Furthermore, the authors propose that this approach could aid in predicting outcomes for Beam Energy Scan (BES) studies, provided that data at specific energies and reliable theoretical constraints are available.