Machine Learning Methods for Stellar Collisions. I. Predicting Outcomes of SPH Simulations

Technical Summary: Machine Learning Methods for Stellar Collisions. I. Predicting Outcomes of SPH Simulations

Problem Statement
Stellar collisions are frequent in dense environments such as galactic nuclei and globular clusters, playing a critical role in the formation of exotic objects (e.g., blue stragglers, massive black hole seeds) and transient events. While Smoothed Particle Hydrodynamics (SPH) simulations provide accurate predictions of collision outcomes (mergers, disruptions, or flybys) and remnant properties, their computational cost renders them infeasible for "on-the-fly" integration into large-scale $N$-body simulations of star cluster dynamics. Existing analytic fitting formulae often fail to generalize across different energy scales or stellar evolutionary phases. Consequently, there is a need for a computationally efficient method to predict collision outcomes that retains the physical fidelity of hydrodynamic simulations.

Methodology
The authors present a comprehensive framework combining high-resolution SPH simulations with machine learning (ML) techniques.

1. Simulation Grid Generation:

   • A new grid of 27,720 SPH simulations was computed using the StarSmasher code.
   • Initial Conditions: Stars were modeled using MESA (Modules for Experiments in Stellar Astrophysics) profiles with a metallicity of $Z = 0.01 Z_{\odot}$. Only main-sequence (MS) stars were considered, with ages sampled logarithmically from $10^{-3}$ to $13.7$ Gyr.
   • Parameter Space: The grid spans primary and secondary masses ($0.2 - 64 M_{\odot}$), velocities at infinity ($10 - 16,000$ km/s), and pericenter distances (sampled based on fractional enclosed mass).
   • Post-processing: Bound masses were determined using an iterative energy-based procedure. Simulations were terminated dynamically based on orbital stability and thermal readjustment.

2. Machine Learning Framework:

   • Input Features: Five parameters were transformed to comparable ranges: $\log(t+0.001)$, $\log(r_p+0.1)$, $\log(v_{\infty}+10)$, $\ln(M_1)$, and $\ln(M_2)$. Data were standardized using mean and standard deviation.
   • Classification Task: The goal was to predict the qualitative outcome into four classes:
     • Class 0: Mutual destruction (0 survivors).
     • Class 1: Merger (1 survivor).
     • Class 2: Flyby (2 survivors).
     • Class 3: Stripping (1 survivor, but distinct from a merger due to envelope loss).
   • Regression Task: The goal was to predict the final mass fractions of the surviving components ($M_{1,f}/M_{tot,i}$, $M_{2,f}/M_{tot,i}$, and unbound mass fraction), enforcing mass conservation.
   • Algorithms Compared: The study evaluated $k$-Nearest Neighbors ($k$-NN), Support Vector Machines (SVM/SVC), and Neural Networks (NN).
   • Handling Class Imbalance: The dataset exhibited significant imbalance (e.g., stripping and mutual destruction were minority classes). The authors utilized balanced accuracy as the primary metric and applied class weighting in the loss functions for SVM and NN to penalize misclassifications of minority classes. Data augmentation (e.g., SMOTE) was explicitly avoided to prevent introducing synthetic artifacts.
   • Mixture of Experts (MoE): A specialized architecture was tested where a shared backbone feeds into a classification head, which acts as a gating mechanism to route data to class-specific regression experts.

Key Results

• Classification Performance:

  • The Neural Network (NN) achieved the highest balanced accuracy of 98.4%, outperforming the Support Vector Classifier (SVC) at 97.7% and $k$-NN at 91.7%.
  • The NN demonstrated superior performance on minority classes (particularly Class 3, stripping), attributed to the use of weighted cross-entropy loss.
  • Decision boundary analysis revealed that the NN produced the most physically consistent and smooth boundaries, whereas $k$-NN showed artificial features and SVC exhibited discontinuities in high-dimensional projections.

• Regression Performance:

  • The NN significantly outperformed both $k$-NN and Support Vector Regression (SVR) in predicting final masses.
  • Relative Errors: The NN achieved median relative errors as low as 0.11% for the final star 1 mass and 0.15% for star 2 mass.
  • Absolute Errors: The median absolute error for $M_{1,f}$ was $4.67 \times 10^{-3} M_{\odot}$ (best run: $4.41 \times 10^{-3} M_{\odot}$), and for $M_{2,f}$ was $4.7 \times 10^{-4} M_{\odot}$.
  • SVR struggled with $M_{2,f}$, exhibiting errors approximately three orders of magnitude larger than $k$-NN in some cases, though relative errors remained below 2% for surviving stars.

• Mixture of Experts (MoE):

  • The MoE architecture achieved a classification balanced accuracy of 98.5%, comparable to the standalone NN.
  • However, for regression tasks, the individually trained NN outperformed the MoE across all ground-truth classes. The MoE showed constrained error distributions for specific classes (reducing high-error outliers) but generally yielded higher mean and median errors than the specialized single-task NNs.

Significance and Contributions
The paper claims that the primary contribution is the creation of a publicly available software package, collAIder, which encapsulates the trained ML models. This tool enables rapid, accurate predictions of stellar collision outcomes within $N$-body simulations of dense star clusters, a task previously limited by the prohibitive cost of running SPH simulations on-the-fly.

The authors emphasize that their approach:

1. Generalizes across scales: By training on a grid spanning a wide range of masses, ages, and velocities, the models overcome the limitations of previous analytic formulae that were often restricted to specific regimes.
2. Improves Physical Fidelity: The ML models capture complex hydrodynamic dependencies (such as the transition between merger and stripping) that simple "sticky sphere" approximations in $N$-body codes miss.
3. Validates ML Utility: The study demonstrates that deep learning, specifically Neural Networks with appropriate loss weighting, is superior to traditional ML methods for this high-dimensional, non-linear astrophysical problem.

The authors remain modest regarding future implications, noting that while the current grid is restricted to main-sequence stars, future work will expand to pre-main-sequence and giant star encounters, as well as incorporating stellar rotation and metallicity-dependent winds. They also suggest that active learning could be used in the future to optimize dataset expansion.