Alex Khan, BosonQ Psi Corp; Dhamotharan S, BosonQ Psi Corp; Divyansh Deshmukh, BosonQ Psi Corp; Aditi Lal, BosonQ Psi Corp; Jay Shah, BosonQ Psi Corp; Parth Danve, BosonQ Psi Corp; Rut Lineswala, BosonQ Psi Corp; Abhishek Chopra, BosonQ Psi Corp
https://doi.org/10.64861/BTAT2567
Keywords: SDA, QAPINN, PINN, Propagation, AI, Edge, Autonomy, Orekit, Maneuver, Surrogate, Conjunction, Uncertainty
Abstract:
Real-time autonomous satellite operations, including conjunction assessment, collision avoidance, and rendezvous and proximity operations, are limited by onboard compute. Analytical propagators such as SGP4 are fast but imprecise, while high-fidelity numerical libraries such as Orekit are accurate but too demanding for typical flight computers, increasing ground-station dependencies and limiting onboard decision making. We present a resource-efficient propagation framework built on physics-informed machine learning. A physics-informed neural network (PINN) embeds the equations of motion in its loss function, but because these terms act only as soft penalties, we add hard constraints so predictions stay physically consistent. Our models achieve 0.7 km and 0.18 km of average error for a propagation of 40,000 seconds for LEO and GEO satellites, respectively.
We also investigate a hybrid Quantum-Assisted PINN (QA-PINN) that replaces the classical hidden layer with a variational quantum circuit, a parameterization shown in fluid-mechanics benchmarks to reduce trainable parameters without loss of accuracy. The QA-PINN model converges faster at only 20 epochs with deviation error 64% less than that of the classical model. On an NVIDIA Jetson Orin Nano, we benchmark a hard-constrained PINN and QA-PINN, comparing agreement with reference dynamics, throughput, and memory footprint. A comparative analysis of execution time on CPU and GPU (Tesla T4) between Orekit, SGP4, PINN, and QA-PINN is discussed.
The proposed propagators are used to study maneuver detection and conjunction analysis. We further integrate uncertainty quantification to study aleatoric and epistemic uncertainties associated with the propagators. Together, these results provide a systems-level assessment of physics-informed and quantum-assisted propagation for onboard, real-time orbit prediction under strict compute budgets.
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications