ML in Astrophysical Turbulence I: Predicting Prestellar Cores in Magnetized Molecular Clouds using eXtreme Gradient Boosting

The gist: This work presents a supervised machine learning framework to forecast the Lagrangian history of prestellar cores in magnetohydrodynamic (MHD) turbulence by mapping instantaneous phase-space states to future coordinates over a horizon of approximately 0.45 Myr.

Simulation Setup and Data Generation

The researchers utilized the open-source adaptive mesh refinement (AMR) code Enzo with the constrained transport magnetohydrodynamics (MHD) module to model the turbulent environment. Initial conditions were generated following the procedure of Mac Low, which involved initializing a periodic turbulent box with uniform density and magnetic field, injecting kinetic energy at large scales to cascade it down, and then activating self-gravity. To track gas parcels in a Lagrangian framework, passive tracer particles were included that followed the flow via a piecewise linear interpolation of the grid velocity field. The simulation evolved for approximately one free-fall time, tff = p/3π/32Gρ0, and 1283 tracer particles were initialized to yield a total of 2,097,152 particles.

Data Preparation for Machine Learning

The dataset MCDS1 was constructed by pairing the input features at snapshot i with the target coordinates at snapshot i + ∆frame where X =

ID, Frame, x, y, z, vx, vy, vz, and log10 ρ>. The prediction horizon was set to a lag of ∆frame = 30 snapshots. This choice corresponds to ∆T ≈ 0.45 Myr and was selected to match the integral time scale of the turbulence. Core particles were identified by employing a dendrogram algorithm on the final simulation snapshot, tagging all tracer particles located within the densest zones of these peaks as ”core particles”. This selection procedure isolated ∼ 117,364 particles (∼ 5.6% of the total population) that successfully accrete onto prestellar cores.

Model Architecture and Training

Extreme Gradient Boosting (XGBoost) was employed as the regression model to approximate the mapping function Φ: R n → R cubed, where the objective is to minimize the expected loss L(yi, Φ(xi)). The model approximates the target variable by summing the outputs of K regression trees: yˆi = Φ(xi) = X K k=1 fk(xi). Tree-based models were selected because they naturally handle sharp discontinuities present in supersonic turbulent flows (shocks), and they are invariant to the monotonic scaling of features, requiring less preprocessing of variables like density. Hyperparameter optimization involved tuning parameters such as n estimators ∈ [500, 1500], eta (Learning Rate) ∈ [0.01, 0.3], and gamma ∈ [0.5, 5] via repeated k-fold cross-validation.

Performance Evaluation and Validation

The model's performance was evaluated using four statistical metrics: Periodic Mean Absolute Error (P-MAE), Periodic Coefficient of Determination (R2p), Cross-Validation Uncertainty (σCV), and the Bounded Accuracy Fraction (Ak). Model A, the optimized configuration, achieved a global MAE of 0.0314 pc with R2 = 0.9988. Temporal generalization tests showed that the model significantly outperformed the No-Motion Baseline across all future frames. Furthermore, testing on an independent simulation run (u502) with a different magnetic field strength demonstrated that the model retained predictive power, showing a 50 − 60% improvement over the baseline.

Physical Implications and Future Directions

The study confirmed that the instantaneous phase-space state (position, velocity, and density) contains sufficient information to predict short-term dynamical evolution with high accuracy (R2 > 0.99). The inclusion of velocity information was critical for distinguishing between transient fluctuations and true gravitationally bound cores. The model's success in predicting the convergent flow patterns of star-forming regions confirms that the algorithm can be used as a reliable ”lagrangian tracker” to identify the provenance of star-forming gas. Future work aims to incorporate spatial correlations by leveraging Deep Learning architectures, specifically 3D Convolutional Neural Networks (CNNs) and Residual Networks (ResNets), to ingest the full volumetric simulation grid directly. This combination of Lagrangian tracking and spatial context awareness is intended to build a unified predictive model for star formation.

Limitations

Despite its success, the model has inherent limitations common to data-driven approximations, including error propagation as recursive application leads to divergence consistent with the chaotic nature of turbulence. Furthermore, XGBoost does not enforce conservation of mass, momentum, or energy. The current model relies on full 3D state vectors (x, y, z, vx, vy, vz), which are accessible in simulations but not in observations. Finally, the decision to exclude B-field features was based on preliminary analyses indicating that the instantaneous local magnetic field strength possessed a relatively low correlation with the short-term future trajectory of the gas parcels. The paper concludes that this approach represents a ”first-order” method, utilizing only local, point-wise information.

Conclusion

The researchers successfully trained an Extreme Gradient Boosting (XGBoost) model to forecast the 3D positions of gas parcels over a horizon of ∼ 0.45 Myr (0.25 tff). The key findings are that local phase-space information (position, velocity, and density) contains sufficient predictive power regarding the future evolution of the gas. The model successfully distinguished between transient density fluctuations and gravitationally bound collapsing cores, achieving a high coefficient of determination (R2 > 0.99). The tree-based ensemble proved to be computationally efficient and capable of handling the high dynamic range of astrophysical data without the extensive normalization required by neural networks. This work represents a ”first-order” approach, utilizing only local, point-wise information. The natural extension is to incorporate spatial correlations by leveraging Deep Learning architectures, specifically 3D Convolutional Neural Networks (CNNs) and Residual Networks (ResNets).

Acknowledgements

Support for this work was provided in part by the National Science Foundation under Grant AAG1616026. This work used Anvil at Purdue University (Song et al. 2022) through allocation PHY240036 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by U.S. National Science Foundation grants.

References

The paper cites numerous foundational works including Larson 1981, McKee & Ostriker 2007, and Chen & Guestrin 2016. The references listed include works by Alatalo et al. 2013, Bate et al. 1995, and Collins et al. 2012 <ref:2603.

1. Improved AI System Can Predict Lagrangian History of Prestellar Cores using XGBoost by mapping instantaneous phase-space state (position, velocity, and density) to future coordinates over a horizon of approximately 0.45 Myr (0.25 tff).

2. The improved system can distinguish between transient density fluctuations and bound collapsing cores by utilizing the local phase-space information alone as shown in the paper, achieving a global coefficient of determination of R2 > 0.99 and capturing the non-linear convergent flows characteristic of gravitational collapse.

3. The system can provide a computationally efficient alternative to traditional sink-particle algorithms by transforming the Lagrangian history problem into a tabular regression task using XGBoost, which is highly effective for tabular physical data and offers interpretability regarding feature importance.

4. The system can be implemented on-the-fly within lower-resolution simulations by sampling tracer particles and passing their state vectors through the trained XGBoost model to statistically predict the Star Formation Rate (SFR) and sink particle locations without explicitly resolving the computationally expensive collapse phase.

5. The system can reconstruct the formation history of prestellar cores by recursively predicting their trajectories backward and forward in time, with Model A achieving a Bounded Accuracy of A3 approximately 91% for core particle predictions.

Giant Molecular Clouds (GMCs) are dominated by supersonic turbulence and magnetic fields, creating a complex network of shocks and filaments that regulate star formation, making it challenging to predict which individual gas parcels will collapse into star-forming cores. Here, we present an XGBoost machine learning framework to forecast the 3D Lagrangian trajectories and gravitational collapse fate of prestellar gas. Models were trained on about 2.1 million tracer particles from self-gravitating, turbulent magnetohydrodynamic (MHD) simulations. By mapping instantaneous point-wise states (position, velocity, density, and magnetic fields) to future coordinates, our regression model predicts continuous 3D spatial evolution over 0.25 t ff (about 0.31 Myr) with R squared > 0.99. Concurrently, our classifier accurately ranks gravitational core susceptibility (AUC = 0.975). Using SHapley Additive exPlanations (SHAP), we interpret the model's decision hierarchy, revealing that local density, kinetic energy, and magnetic field magnitude dominate core identification. SHAP dependence analysis shows that the model re-derives a sharp non-linear collapse threshold at 10ρ about 0.8 above which magnetic fields amplify collapse predictions. The classification task reveals a conservative false-positive rate (37% precision) at absolute binary thresholds. We demonstrate that while point-wise ML models excel at kinematic advection, they mathematically lack the spatial context required to evaluate local differential operators (such as magnetic directional derivatives, thermal pressure gradients, or field curvature) that are physically necessary to stop collapse. This data-driven approach offers a computationally efficient foundation for advective subgrid modeling while highlighting the necessity of spatial architectures to establish absolute core-formation criteria.