PINN Architecture for Satellite Gravimetry

Physics-Informed Loss Function

\[\mathcal{L}_{total} = \mathcal{L}_{data} + \lambda_{physics}\mathcal{L}_{physics} + \lambda_{BC}\mathcal{L}_{BC}\]

Where physics loss enforces Laplace's equation: \(\nabla^2 V = 0\) in free space

4 layers
64 neurons
λ = 0.10
--
Total Parameters
3
Input Dimensions (r,θ,φ)
4
Output (V, ∂V/∂r, ∂V/∂θ, ∂V/∂φ)
--
Memory (MB)

Training Dynamics & Loss Convergence

1e-3
5000 points
Ready to train
0 / 1000 epochs
0
Current Epoch
--
Total Loss
--
Physics Loss
--
Data Loss

Gravity Field Reconstruction

Spherical Harmonic Expansion

\[V(r,\theta,\phi) = \frac{GM}{r}\sum_{n=0}^{N}\sum_{m=0}^{n}\left(\frac{R}{r}\right)^n P_{nm}(\cos\theta)[C_{nm}\cos(m\phi) + S_{nm}\sin(m\phi)]\]

N = 60
450 km
50%
--
RMS Error (mGal)
--
Max Error (mGal)
--
Coefficients Recovered
--
Spatial Resolution (km)

PINN vs Traditional Methods

2.0 mGal
20%

PINN Advantages

• Physics constraints reduce required training data by 60-80%
• Automatic differentiation ensures exact derivative computation
• Handles irregular data distributions naturally
• Implicit regularization from physics loss

Traditional Method Strengths

• Established error propagation theory
• Faster for simple linear problems
• Well-understood convergence properties
• Mature software implementations

Hybrid Approach

• Use traditional methods for initial solution
• Apply PINN for gap-filling and refinement
• Combine uncertainties using ensemble methods
• Best of both worlds for operational use