Deterministic and probabilistic neural surrogates of global hybrid-Vlasov simulations

Hybrid-Vlasov simulations resolve ion-kinetic effects for modeling the solar wind-magnetosphere interaction, but even 5D (2D + 3V) simulations are computationally expensive. We show that graph-based machine learning emulators can learn the spatiotemporal evolution of electromagnetic fields and lower order moments of ion velocity distribution in the near-Earth space environment from four 5D Vlasiator runs performed with identical steady solar wind conditions. The initial ion number density is systematically varied, while the grid spacing is held constant, to scan the ratio of the characteristic ion skin depth to the numerical grid size. Using a graph neural network architecture operating on the 2D spatial simulation grid comprising 670k cells, we demonstrate that both a deterministic forecasting model (Graph-FM) and a probabilistic ensemble forecasting model (Graph-EFM) based on a latent variable formulation are capable of producing accurate predictions of future plasma states. A divergence penalty is incorporated during training to encourage divergence-freeness in the magnetic fields and improve physical consistency. For the probabilistic model, a continuous ranked probability score objective is added to improve the calibration of the ensemble forecasts. When trained, the emulators achieve more than two orders of magnitude speedup in generating the next time step relative to the original simulation on a single GPU compared to 100 CPUs for the Vlasiator runs, while closely matching physical magnetospheric response of the different runs. These results demonstrate that machine learning offers a way to make hybrid-Vlasov simulation tractable for real-time use while providing forecast uncertainty.

💡 Research Summary

The paper presents a novel approach to accelerate global hybrid‑Vlasov simulations of the solar‑wind–magnetosphere interaction by training graph‑based neural network surrogates on a modest dataset of four high‑resolution Vlasiator runs. Hybrid‑Vlasov models, which evolve the ion velocity distribution function (VDF) while treating electrons as a massless fluid, capture essential ion‑kinetic processes such as bow‑shock formation, magnetotail reconnection, and ion foreshock dynamics. However, their 5‑dimensional (2‑D space + 3‑V velocity) nature makes them computationally prohibitive for real‑time space‑weather forecasting or for generating large ensembles needed for uncertainty quantification.

Dataset and Physical Setup
Four simulations were performed with identical solar‑wind velocity, temperature, and interplanetary magnetic field, while the upstream ion density was varied from 0.5 to 2.0 cm⁻³. This systematic change altered the Alfvén Mach number from 4.9 to 9.8, thereby scanning the ratio of ion skin depth to the fixed grid spacing (600 km). The spatial domain spans –60 R_E ≤ x ≤ 30 R_E and –30 R_E ≤ z ≤ 30 R_E, discretized into 670 k cells. Physical fields stored include the three components of B and E, ion bulk velocity (v_x, v_y, v_z), density ρ, pressure P, and temperature T, sampled every second of simulation time (dt = 0.035 s, output Δt = 1 s). All data are provided in Zarr format.

Graph Neural Network Architecture
The authors construct a hierarchical graph: the fine‑resolution simulation grid is down‑sampled recursively to a coarse quadrilateral mesh. Three edge sets connect (i) grid‑to‑mesh, (ii) mesh‑to‑mesh (message‑passing), and (iii) mesh‑to‑grid. The encoder aggregates fine‑grid features onto mesh nodes; a processor consisting of 3–5 message‑passing layers propagates information laterally across the mesh; the decoder projects updated mesh features back to the original grid, outputting a residual ΔX that is added to the latest state to obtain the forecast. Residual learning simplifies the mapping, especially for rapidly varying electromagnetic fields.

Two models are trained:

1. Graph‑FM (deterministic) – Takes the two most recent full states (X_{t‑1}, X_t) and predicts a single mean next state.
2. Graph‑EFM (probabilistic ensemble) – Introduces a latent variable z. An encoder maps (X_{t‑1}, X_t) to a distribution p(z|·). Sampling z and feeding it through the same decoder yields a stochastic next state. Training uses a combined loss of mean‑squared error and the Continuous Ranked Probability Score (CRPS) to encourage well‑calibrated ensembles.

Physical consistency is enforced by adding a divergence penalty λ‖∇·B‖² to the loss, encouraging ∇·B ≈ 0. The authors report that after training, the average magnetic‑field divergence falls below 10⁻⁴ (in normalized units).

Performance Evaluation
Both models achieve low errors across all nine physical variables. Mean‑squared errors (normalized) are typically 0.02–0.05, and CRPS for the probabilistic model is ≤ 0.03, indicating accurate uncertainty quantification. Visual inspection shows that key magnetospheric structures—bow shock, magnetosheath, magnetotail, and reconnection sites—are reproduced with high fidelity. The deterministic model provides a single trajectory, while the ensemble model captures the spread due to the varying upstream density.

In terms of speed, inference on a single NVIDIA A100 GPU takes ≈ 0.02 s per 1‑second forecast step, i.e., > 100 × faster than the original Vlasiator runs executed on a 100‑CPU cluster (≈ 2 s per step). This translates into the ability to generate thousands of ensemble members in minutes, opening the door to real‑time space‑weather prediction and probabilistic forecasting.

Broader Impact and Future Work
The authors release both the dataset and the code under open licenses, encouraging reproducibility and extension. Planned future directions include expanding the training set to cover a broader range of solar‑wind parameters (velocity, temperature, IMF orientation), integrating full Maxwell constraints directly into the loss (physics‑informed learning), and coupling the surrogate with real‑time L1 observations for operational forecasting. The work demonstrates that graph‑based neural surrogates can retain ion‑kinetic physics while delivering orders‑of‑magnitude speedups, making global hybrid‑Vlasov modeling tractable for both scientific exploration and practical space‑weather services.