Physics-Informed Neural Networks for Modeling the Martian Induced Magnetosphere

Jiawei Gao, Chuanfei Dong, Chi Zhang, Yilan Qin, Simin Shekarpaz, Xinmin Li, Liang Wang, Hongyang Zhou, Abigail TadlockarXiv 2025 · DOI · arXiv

Introduction

Modeling the complex magnetic environment of Mars represents a significant challenge in planetary science. Unlike Earth, Mars lacks a global intrinsic magnetic field. Instead, the interaction between the solar wind—a stream of charged particles from the Sun—and the Martian upper atmosphere creates an induced magnetosphere. This environment is critical to understand, as it governs the rate at which Mars loses its atmosphere to space.

The Modeling Dilemma

Historically, researchers have had to choose between two imperfect modeling approaches:

  1. Physics-based Magnetohydrodynamic (MHD) simulations: These are highly accurate but computationally expensive, requiring significant time to simulate specific conditions.
  2. Statistical Empirical Models: These are fast but often rely on simple averaging of data, which can obscure the dynamic response of the magnetosphere to varying solar wind conditions.

To bridge this gap, researchers have developed a 3D model using Physics-Informed Neural Networks (PINNs), combining the speed of data-driven methods with the rigor of physical laws.

Methods: Integrating Physics into Deep Learning

The core innovation of this study is the PINN architecture. Unlike standard neural networks that act as "black boxes," a PINN incorporates fundamental physical constraints directly into its learning process via a specialized loss function.

The PINN architecture and its associated loss functions, including data, physics, and boundary condition constraints.

The model consists of a fully connected feed-forward network with three hidden layers of 128 neurons each. The training process is constrained by:

  • Gauss's Law for Magnetism: The model is forced to satisfy $\nabla \cdot \mathbf{B} = 0$, ensuring magnetic field lines are continuous.
  • Boundary Conditions: Penalties are applied to ensure the field matches the Interplanetary Magnetic Field (IMF) at large distances and remains horizontal at the Martian surface.

Datasets and Drivers

The model was trained using eight years of data (2014–2022) from the MAVEN (Mars Atmosphere and Volatile EvolutioN) spacecraft. The dataset includes approximately 1.61 million filtered data points.

Spatial distribution of MAVEN data points in MSE coordinates and histograms of the solar wind drivers: dynamic pressure, IMF intensity, and cone angle.

The network takes six inputs: the spacecraft's 3D position $(x, y, z)$ and three solar wind drivers:

  • Dynamic Pressure ($P_{SW}$): Influences the compression of the magnetic barrier.
  • IMF Intensity ($B_{IMF}$): The primary driver of induced field strength.
  • IMF Cone Angle ($\theta_{cone}$): Affects the orientation and "kinking" of the field lines.

Key Results and Insights

The PINN-based approach yielded several significant findings:

  • Physical Consistency: The inclusion of physics as a regularizer prevented overfitting and ensured that field lines did not "bleed" into the planet, a common error in purely data-driven models.
  • Structural Reconstruction: The model successfully captured complex features like the Inverse Polarity Reversal Layer (IPRL) and the asymmetric "draping" of magnetic fields, which is stronger in the hemisphere aligned with the solar wind electric field. The best model (PINN-A1) achieved a root-mean-square (RMS) error of approximately 4.23 nT, showing excellent agreement with actual MAVEN observations.
  • Efficiency: Once trained, the PINN can reconstruct the entire 3D magnetosphere for any given solar wind condition almost instantaneously, a task that would take hours or days for traditional MHD simulations.

Limitations and Future Outlook

While a major step forward, the current model has limitations. It excludes Mars' localized crustal magnetic fields, which are significant in the southern hemisphere. Future iterations aim to include these crustal fields, incorporate additional drivers like solar UV flux, and utilize multi-satellite data from upcoming missions like EscaPADE to further refine our understanding of the Martian environment.

Summary updated

BibTeX
@misc{gao2025physics,
  author = {Jiawei Gao and Chuanfei Dong and Chi Zhang and Yilan Qin and Simin Shekarpaz and Xinmin Li and Liang Wang and Hongyang Zhou and Abigail Tadlock},
  title = {{Physics-Informed Neural Networks for Modeling the Martian Induced Magnetosphere}},
  year = {2025},
  eprint = {2512.16175},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.16175},
}
Figure