Machine‑learning‑based shape estimation of deployable RF antennas using temperature data

Abstract

Thermally induced surface deformation can degrade the performance of large deployable radio-frequency antennas, motivating methods for in situ shape estimation. Previous work demonstrated that a feed-forward neural network (NN) could predict panel deformation from sparse temperature measurements; however, the accuracy of a purely data-driven model depends on the quantity and distribution of its training data. This work introduces a physics-informed residual model that combines a data-driven reduced-order thermoelastic approximation based on proper orthogonal decomposition (POD) with a shallow NN correction. The POD-residual model was compared with a direct NN under nominal interpolation, limited-training-data, and magnitude-extrapolation conditions. Under limited training data, the residual model reduced the mean maximum root-mean-square error by approximately 50% and maintained prediction errors below 0.17mm across all five test cases. Under magnitude extrapolation, both models retained bounded errors without catastrophic growth. These results demonstrate that the POD-residual formulation improves robustness to reduced training data while maintaining useful prediction stability beyond the represented deformation-amplitude range.