Application of neural networks for real-time erosion monitoring and adaptive attitude/morphology control in low earth orbit/very low earth orbit

• Physics-Informed Neural Networks (PINN) integrate the principles of orbital mechanics and erosion physics to enable autonomous satellite damage mitigation. • Nano-smart materials: state-of-the-art technology for selecting the most suitable material for the system. • Algorithm structure: PINN, mission parameters, dataset generation and normalization • Predictive analysis and evaluation of training and test accuracy. Satellites operating in Low Earth Orbit (LEO) and Very Low Earth Orbit (VLEO) are subject to extreme environmental stressors, including high energy radiation, significant thermal gradients, atomic oxygen (AO) erosion, and frequent impacts from orbital debris. These conditions pose substantial challenges to structural integrity and mission longevity. To increase robustness and extend operational life, smart materials equipped with embedded sensors and self-actuating mechanisms capable of real-time structural adaptability are integrated. This study presents a model that is parameterized using the mission-specific characteristics of a 10U CubeSat configuration, leveraging a Physics Informed Neural Network (PINN). The model incorporates inertial properties, control authority constraints and environmental conditions to predict attitude dynamics feedback and trigger adaptive responses based on micrometroids-induced erosion or penetration. The system monitors key parameters, such as temperature fluctuations, surface erosion, and penetration events transmitting sensor data to an onboard processor. A deep neural network interprets this information to trigger corrective actions: either through dynamic shape morphing or attitude control maneuvers. Coupled quaternion-based kinematics, rigid body rotational dynamics, and AO driven surface degradation are simulated in a custom virtual environment to generate synthetic, noise corrupted datasets. These data pairs are used to train a feedforward PINN, which maps erosion states to discrete rotational deviations and surface morphology adaptation triggers. Unlike traditional black box models, the PINN is grounded in physical laws, enabling real-time, onboard inference of erosion induced anomalies with improved robustness. While achieving robust average performance, capturing the full complexity of thresholded, rare impulse events remain challenging, and reliance on synthetic data alone cannot encompass every in-orbit phenomenon. Nevertheless, grounding the inference in first principles allows the network to generalize more reliably when confronted with novel erosion scenarios. The lightweight surrogate architecture is optimized for flight representative hardware, supporting fully autonomous anomaly detection and response without ground intervention and the integration of sensor driven morphology adaptation with conventional control strategies broadens the design space for resilient small satellites. Moreover, the framework is inherently extensible: it can be applied to more complex geometries and multi-axis erosion effects, and it readily accommodates hybrid training that incorporates on-ground test data or live in-orbit telemetry.