Physics-informed neural networks for spacecraft trajectory optimization with exact boundary conditions

Samuel Schönherr, Günther Waxenegger-Wilfing, Sascha RanftlarXiv 2026 · arXiv

TL;DR

This paper introduces a method to enforce exact boundary conditions in Physics-Informed Neural Networks (PINNs) for spacecraft trajectory optimization. By mathematically embedding initial and terminal states directly into the network's architecture, the authors eliminate the need for sensitive loss-weight tuning, leading to more stable convergence and higher accuracy in complex orbital maneuvers.

What Problem Does This Paper Address?

Designing a spacecraft's path—trajectory optimization—is a fundamental challenge in aerospace engineering. Recently, Physics-Informed Neural Networks (PINNs) have emerged as a powerful tool for this, as they can solve the underlying differential equations of motion in a continuous, grid-free way.

However, "standard" PINNs face a major hurdle: they treat boundary conditions (like starting at a specific point on Earth and ending at a specific orbit around Mars) as penalty terms in a multi-objective loss function. This creates a difficult "balancing act." If the weight for the boundary loss is too low, the spacecraft misses its target; if it's too high, the network ignores the physics of gravity and thrust. This sensitivity often leads to poor convergence or total failure in complex missions.

Methods

The authors propose a reparameterization principle to move from soft constraints to hard constraints.

  • Exact BC Transformation: Instead of using a raw neural network to predict position, they apply a mathematical transformation that incorporates "distance functions." These functions are designed to be zero at the start and end times, effectively "pinning" the trajectory to the required boundary conditions regardless of the network's internal weights.
  • Single-Objective Optimization: Because the boundary conditions are satisfied by construction, the training process no longer needs to balance competing loss terms. It becomes a single-objective problem: minimizing the total thrust (control effort) while satisfying the laws of physics.
  • Learnable Time-of-Flight (TOF): The authors treat the mission duration as an optimizable parameter. Because the entire system is auto-differentiable, the network can learn the optimal time required to complete a maneuver alongside the path itself.
  • Coordinate Flexibility: The method is demonstrated in both Cartesian and polar coordinates, showing it can adapt to different mission geometries (like circular orbit transfers).

Datasets

The paper does not use external datasets. Instead, it evaluates the method on five synthetic but physically rigorous benchmark problems of increasing complexity:

  1. 2D Swing-by: Navigating past three celestial bodies in a plane.
  2. 3D Swing-by: The same maneuver but allowing out-of-plane movement.
  3. Circular Orbit Transfer (Fixed Angle): Moving from Low Earth Orbit (LEO) to Medium Earth Orbit (MEO).
  4. Circular Orbit Transfer (Free Angle): Moving from LEO to Geostationary Earth Orbit (GEO) where the final position on the orbit is flexible.
  5. Hold-Point Rendezvous: A proximity operations scenario in LEO.

Results are compared against OpenGoddard, a classical numerical trajectory optimization solver.

Key Insights and Ideas

  • Exploiting Boundary Structure: Trajectory optimization is unique because its boundaries are "low-dimensional"—usually just two points in time (start and end). The authors realized this makes it much easier to build stable distance functions compared to the complex 3D boundaries found in general fluid dynamics or structural mechanics PDEs.
  • Eliminating Gradient Pathologies: By removing the boundary loss terms, the authors eliminate "gradient pathologies"—situations where the gradients from different loss terms point in opposite directions, causing the optimizer to stall.
  • Bifurcations and Local Minima: The authors discovered that the Time-of-Flight creates a complex "loss landscape." In some missions, there are two distinct valid solutions (e.g., a direct path vs. a long detour). They found that "pre-conditioning" the network with auxiliary velocity constraints can help guide the optimizer toward the desired family of solutions.

Key Results

  • Superior Convergence: In the 2D swing-by test, the transformed PINN achieved a mean thrust residual ($\Delta V$) of $2.06 \times 10^{-3}$, significantly outperforming the "naive" PINN's $5.84 \times 10^{-3}$.
  • Enabling Complex Missions: For the more advanced orbital transfers (LEO to MEO/GEO), the standard "naive" PINN failed to converge entirely. The transformed PINN, however, successfully found smooth, physically meaningful low-thrust trajectories.
  • Speed and Stability: The proposed method was generally faster to train (e.g., 23.5s vs 30.7s for the 2D case) because the optimizer didn't have to "fight" to satisfy the boundary conditions.
  • Comparison to Classical Solvers: While classical solvers like OpenGoddard still found slightly more optimal solutions (lower total $\Delta V$), the PINN approach provides a continuous, differentiable representation that is highly valuable for sensitivity analysis and real-time guidance applications.

Limitations and Future Work

  • Non-Convexity: The optimization landscape still contains multiple local minima. In higher-dimensional settings like the 3D swing-by, the method still requires smart initialization or pre-conditioning to find the global optimum.
  • Inequality Constraints: The current framework handles "equality" constraints (must be at point X). It does not yet incorporate "inequality" constraints, such as maximum thrust limits or avoiding "no-go" zones in space.
  • Long-Duration Missions: For very long missions, a single neural network might struggle to capture all the dynamics. The authors suggest that combining this method with "domain decomposition" (breaking the mission into smaller time segments) is a promising next step.

Summary updated

Abstract

We use neural networks to optimize trajectories by modeling the position as a function of time and minimizing the thrust with respect to the network parameters. This corresponds to an unsupervised training procedure in which physical laws are incorporated directly into the optimization objective. In standard physics-informed neural networks, initial and terminal states are enforced through additional weighted loss terms, resulting in a multi-objective optimization problem that is highly sensitive to the choice of loss weights and can lead to poor convergence and reduced accuracy. To address this challenge, we construct a simple transformation, exploiting the mathematical structure of trajectory optimization problems, to enforce boundary conditions exactly a-priori. This transformation can be interpreted as a reparameterization of the solution space that embeds the boundary conditions directly into the function. This converts the original constrained optimization problem into an unconstrained one over admissible trajectories. I.e. the transformation eliminates the boundary condition penalty terms and reduces the problem to a single-objective optimization. Crucially, trajectory optimization comes with a low-dimensional boundary, which allows particularly simple and stable constructions of distance functions compared to general domains in partial differential equations. Numerical experiments demonstrate that the proposed approach meaningfully improves convergence behavior and solution accuracy.
BibTeX
@misc{schonherr2026physics,
  author = {Samuel Schönherr and Günther Waxenegger-Wilfing and Sascha Ranftl},
  title = {{Physics-informed neural networks for spacecraft trajectory optimization with exact boundary conditions}},
  year = {2026},
  eprint = {2611.21933},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21933},
}
Figure