Optimizing deep learning models for on-orbit deployment through neural architecture search

Advancements in spaceborne edge computing have facilitated the incorporation of Artificial Intelligence (AI)-powered chips into CubeSats, enabling intelligent data handling and enhanced analytical capabilities with greater operational autonomy. This class of satellites faces stringent energy and memory constraints, necessitating lightweight models typically obtained via compression techniques. This paper addresses model compression through Neural Architecture Search (NAS), enabling computational efficiency and balancing accuracy, size, and latency. More specifically, we design an evolutionary-based NAS framework for onboard processing and evaluate it on both burned-area segmentation and classification tasks. The proposed solution jointly optimizes network architecture and deployment for hardware-specific, resource-constrained platforms, with hardware awareness embedded in the optimization loop to tailor network topologies to the target edge computing chip. The resulting models, designed on CubeSat-class hardware-namely the NVIDIA Jetson AGX [Formula: see text] and Intel® [Formula: see text] Myriad[Formula: see text] X-exhibit a memory footprint below 1 MB, achieving real-time, high-resolution inference in orbit while outperforming handcrafted baselines in terms of latency (3× faster) and maintaining competitive mean Intersection over Union (mIoU). Furthermore, on the classification benchmark conducted on an NVIDIA A100-SXM, our approach attained an atthew Correlation Coefficient (MCC) of 0.974-substantially outperforming the baseline EfficientNet-lite0 (0.902) while achieving a ×47 speedup. These results highlight the framework's scalability, enabling seamless deployment across the spectrum from resource-constrained edge devices to datacenter-grade accelerators, thereby supporting next-generation on-orbit computing architectures.