🔭 astrophysics

Physics-Informed Neural Networks for Modeling the Martian Induced Magnetosphere

This paper introduces the first data-driven model of the Martian induced magnetosphere using Physics-Informed Neural Networks (PINNs) trained on MAVEN observations and physical laws, which successfully reconstructs three-dimensional magnetic field configurations and reveals key dependencies on solar wind conditions.

Original authors: Jiawei Gao, Chuanfei Dong, Chi Zhang, Yilan Qin, Simin Shekarpaz, Xinmin Li, Liang Wang, Hongyang Zhou, Abigail Tadlock

Published 2026-08-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jiawei Gao, Chuanfei Dong, Chi Zhang, Yilan Qin, Simin Shekarpaz, Xinmin Li, Liang Wang, Hongyang Zhou, Abigail Tadlock

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the solar system as a vast, cosmic ocean where a constant stream of charged particles—the solar wind—blows outward from the Sun. When this wind hits a planet, it creates a magnetic "bubble" around it, called a magnetosphere, which acts like a protective shield. For planets like Earth, this shield is built from the planet's own internal magnetic field, much like a giant bar magnet hidden deep inside. But Mars is different; it lost its internal magnet long ago. Instead of a solid shield, Mars has a "induced" magnetosphere. Think of this like a piece of iron floating in a river: the iron doesn't have its own magnetism, but the flowing water (the solar wind) drags and twists the magnetic field lines around it, creating a temporary, squishy shield. Scientists have always wanted to map exactly how this squishy shield behaves, because if they can understand how the solar wind strips away Mars' atmosphere, they can learn how planets lose their ability to support life.

For decades, researchers tried to map this Martian shield using two main tools. The first was like taking a blurry, average photo of a moving crowd; it showed the general shape but missed how the crowd reacted to specific changes. The second was like trying to simulate the entire crowd on a supercomputer; it was incredibly accurate but took so much time and power that it was hard to run many different scenarios. Now, a team of scientists has tried a third, smarter approach. They used a type of artificial intelligence called a Physics-Informed Neural Network (PINN). You can think of this AI as a student who is given a pile of messy, real-world photos taken by a spacecraft, but with a strict rulebook of physics laws (like "magnetic field lines can't just end in empty space") that it must follow while learning. This allows the AI to fill in the gaps and create a smooth, 3D movie of the Martian magnetic shield, showing exactly how it stretches and squeezes under different solar wind conditions.

The team, led by researchers at Boston University, trained this AI using data from the MAVEN spacecraft, which has been orbiting Mars since 2014. They fed the AI real measurements of the magnetic field, along with the "weather" conditions of the solar wind at that time: how hard the wind was pushing (dynamic pressure), how strong the magnetic field in the wind was (IMF intensity), and the angle at which the wind was hitting the planet. The result is a highly detailed, 3D model that successfully reconstructs the invisible magnetic forces surrounding Mars.

The AI discovered some fascinating secrets about how this shield works. First, it confirmed that the strength of the magnetic field wrapping around Mars is mostly controlled by how strong the incoming solar wind's magnetic field is. When the wind's magnetic field is stronger, the shield gets more intense and stretches further out. However, the "push" of the solar wind (its pressure) acts like a hand squeezing a balloon; it doesn't make the magnetic field stronger, but it pushes the whole shield closer to the planet's surface, compressing it.

Perhaps the most playful discovery involves the shape of the shield. The AI found that the magnetic field lines don't just wrap neatly around the planet; they twist into a loop, almost like a hula hoop spinning around Mars. This looping creates a specific magnetic signature that is stronger on one side of the planet than the other, depending on the angle of the solar wind. The model also showed that even when the solar wind hits Mars from extreme angles, this structured shield still forms, though it gets a bit weaker. Interestingly, the AI found that the magnetic field lines in the "tail" of the shield (the part stretching out behind Mars) stay pretty much centered, suggesting that the wind's magnetic field doesn't push this tail off to the side as much as some scientists previously thought.

While the model is a huge step forward, the authors are careful to note its limits. Because the AI was trained to ignore Mars' local, patchy magnetic fields (which are like tiny, stubborn magnets stuck in the planet's crust), the model doesn't show how those specific spots twist the shield. Also, the model is a snapshot of steady conditions and doesn't yet account for the changing seasons on Mars or the varying intensity of the Sun's ultraviolet light. However, by proving that this "physics-aware" AI can learn from sparse data to build a complete picture, the study suggests a new, faster way to understand how solar storms interact with planets. It turns a complex, computationally heavy problem into a flexible tool that could help us better understand not just Mars, but any planet without a strong internal magnetic shield.

Technical Summary: Physics-Informed Neural Networks for Modeling the Martian Induced Magnetosphere

Problem Statement
Mars lacks a global intrinsic magnetic field, resulting in an induced magnetosphere formed by the interaction between the solar wind and the planetary upper atmosphere. This environment is characterized by complex current systems and high variability driven by upstream solar wind conditions. Understanding the magnetic field configuration is critical for studying atmospheric ion escape and evolution. However, existing modeling approaches face significant limitations:

Statistical methods (e.g., slice-averaged maps) provide general configurations but fail to isolate responses to specific solar wind drivers.
Physics-based simulations (e.g., MHD, hybrid) are computationally expensive and rely on underlying model assumptions.
Empirical models based on basis functions (successful for magnetized planets like Earth) are difficult to apply to Mars due to the irregular, spatially variable, and condition-dependent nature of Martian current systems.

Consequently, there is a need for a data-driven approach that can reconstruct continuous, three-dimensional magnetic field structures while adhering to physical laws, utilizing sparse observational data from the Mars Atmosphere and Volatile EvolutioN (MAVEN) mission.

Methodology
The authors developed a data-driven model using Physics-Informed Neural Networks (PINNs) to reconstruct the steady-state three-dimensional magnetic field of the Martian induced magnetosphere.

Data Source: The model was trained on magnetic field and plasma data from MAVEN (November 2014 – May 2022). The dataset was filtered to include only steady Interplanetary Magnetic Field (IMF) conditions and regions where crustal magnetic fields were negligible ( $|B_{crust}| < 10$ nT), resulting in 1.61 million data points.
Coordinate System: Analysis was conducted in Mars–Solar–Electric (MSE) coordinates to account for IMF orientation effects.
Input Parameters: The neural network takes spatial coordinates $(x, y, z)$ $(x, y, z)$ and upstream solar wind parameters as inputs. The study utilized specific combinations:
- PINN-A models: Inputs included solar wind dynamic pressure ( $P_{SW}$ ) and IMF intensity ( $B_{IMF}$ ).
- PINN-B model: Input was the IMF cone angle ( $\theta_{cone}$ ).
Network Architecture: A fully connected feed-forward network with three hidden layers (128 neurons each) using hyperbolic tangent (tanh) activation functions.
Loss Function: The model minimizes a composite loss function ( $L_{total}$ $L_{t o t a l}$ ) comprising:
1. Data Loss ( $L_{data}$ ): Ensures agreement between predicted and observed magnetic fields.
2. Physics Loss ( $L_{phys}$ ): Enforces the divergence-free constraint of the magnetic field ( $\nabla \cdot \mathbf{B} = 0$ ).
3. Boundary Loss ( $L_{bc}$ ): Enforces upstream conditions (matching IMF at $x = 3 R_M$ ) and surface conditions (vanishing normal component at the Martian surface, $R_M$ ).
Ablation Study: Three variations (PINN-A1, A2, A3) were trained to test the necessity of physical and boundary constraints. PINN-A1 (all terms) was identified as the optimal model, avoiding the overfitting of PINN-A3 (data only) and the unphysical field line crossings of PINN-A2 (no boundary loss).

Key Results
The PINN-A1 model successfully reconstructed the global magnetic field structure and its variability under different solar wind conditions:

Global Structure: The model accurately reproduced key features, including IMF draping around the planet, the dayside magnetic barrier, and the extension of draped fields into the magnetotail lobes.
Hemispheric Asymmetry: The model identified asymmetries in the draped field. Draping is stronger in the $+Z_{MSE}$ hemisphere on the dayside, while the $B_Y$ component becomes negative near the southern pole ( $-Z_{MSE}$ ).
Dependence on $B_{IMF}$ : The intensity of the induced magnetic field is primarily controlled by IMF strength. As $B_{IMF}$ increases, field intensity in the magnetosheath and barrier increases, and the spatial extent of draping expands along the $Z_{MSE}$ axis.
Dependence on $P_{SW}$ : Solar wind dynamic pressure governs the compression of the magnetic barrier. Higher $P_{SW}$ compresses the dayside barrier closer to the planet and weakens field intensities near the bow shock, though it has a minor effect on the magnetotail.
Dependence on IMF Cone Angle:
- The model reproduced the Inverse Polarity Reversal Layer (IPRL) in quasi-parallel magnetosheath conditions (cone angles near 0° or 180°), characterized by a local magnetic field minimum near the Magnetic Pileup Boundary (MPB).
- Draping symmetry varies with the cone angle; at 30° (150°), draping is stronger in the $-Y_{MSE}$ ( $+Y_{MSE}$ ) hemisphere, while a 90° angle yields symmetric draping.
- The magnetotail current sheet remains nearly centered regardless of the cone angle, with displacement limited to less than $0.2 R_M$ .
Stability of $-B_Y$ : A significant $-B_Y$ component was observed near the south pole and in the distant magnetotail. This feature remained consistent across all upstream conditions ( $B_{IMF}$ , $P_{SW}$ , and $\theta_{cone}$ ).

Significance and Claims
The paper claims that this work represents the first application of PINNs to model the Martian induced magnetospheric magnetic field. The significance of the study lies in:

Methodological Advancement: Demonstrating that PINNs can effectively integrate sparse, noisy observational data with physical laws (specifically $\nabla \cdot \mathbf{B} = 0$ ) to produce continuous, physically consistent solutions without the computational cost of traditional simulations or the discretization limitations of statistical maps.
Physical Insights: Providing a tool to identify specific dependencies of magnetic field configurations on solar wind drivers, such as the distinct roles of $B_{IMF}$ (magnitude control) and $P_{SW}$ (compression control).
Mechanism Hypothesis: The results support a hypothesis that the observed looping magnetic structures and the persistent $-B_Y$ component arise from a combined effect of plume dynamics (ion pickup and acceleration) and current systems (sunward and tailward currents forming a closed loop).
Future Utility: The approach offers a promising framework for advancing solar wind-Mars interaction studies, particularly as new missions (e.g., EscaPADE, Tianwen-1) provide enhanced real-time monitoring capabilities.

The authors note limitations, including the exclusion of crustal magnetic fields, the use of MSE coordinates (neglecting IMF clock angle), and the omission of other drivers like EUV flux, suggesting these as targets for future modeling iterations.

Technical Summary: Physics-Informed Neural Networks for Modeling the Martian Induced Magnetosphere

More like this