Safety Nets Enable Certifiable AI for Aviation Systems.

Integrating Artificial Intelligence into safety-critical aviation systems poses significant certification hurdles, given the industry's stringent safety requirements. To address this, a "Safety-by-Design" approach using "Safety Nets" has been proposed, which combines neural network compression with lookup tables. This method aims to ensure 100% correct runtime behavior across a discretized operational domain, aligning with EASA guidelines. While Safety Nets have been explored, a comprehensive study on their performance characteristics and design trade-offs was lacking. This work provides the first systematic analysis of the balance between neural network and lookup table sizes within Safety Nets. By comparing various neural network architectures, the study identified optimal design parameters. Architectures with 3 to 5 hidden layers, 50 to 100 nodes per layer, and one-hot encoding proved most effective. These configurations allow neural networks to accurately represent over 97% of the data, with compact lookup tables handling the remaining errors. Crucially, this approach reduces system size by nearly three orders of magnitude, making it compatible with current avionics hardware while guaranteeing complete output correctness. The researchers also released the first open-source implementation for aviation systems like HCAS and VCAS, demonstrating a practical path for certifiable AI.