To AI or not to AI? The AIRGUS Flexible Neural Compression Framework for Scientific Space Data

Recent results indicate that AI-based compression does not always outperform operational CCSDS baselines for scientific space data when evaluated under comparable rate-distortion conditions. This motivates a more selective question: under which data modalities, fidelity constraints, and onboard-resource regimes can AI-based approaches provide credible operational benefit? To answer this question, AIRGUS investigates neural compression for heterogeneous scientific space data, including 1D telemetry, 2D imagery and SAR, and 3D hyperspectral cubes, evaluating neural methods against operationally relevant CCSDS 121/122/123 baselines, under controlled and traceable conditions. The AIRGUS framework flexibly combines predictive AI compression, retaining lossless and near-lossless error-control properties aligned with CCSDS-style coding, with end-to-end neural compression based on learned latent representations. It supports adaptive selection across compression paths according to data modality, rate-distortion constraints, science-fidelity requirements, and onboard resource budgets. The benchmarking roadmap combines classical compression metrics with science-relevant fidelity measures, determinism checks, and deployment KPIs including throughput, memory footprint, and power consumption. The objective is to identify the operational regimes in which neural compression can provide credible advantage over established CCSDS standards, and to define a path toward representative onboard validation.