AI Research

On the Applicability of Safety Nets: A Safety-By-De... | AI Research

Key Takeaways

This paper addresses the challenge of certifying Artificial Intelligence (AI) for use in safety-critical aviation systems.
The integration of Artificial Intelligence (AI) in safety-critical aviation systems presents significant challenges for certification and deployment.
Aviation, often regarded as the safest form of transportation, relies on numerous safety-critical systems.
Although Safety Nets have been studied, no comprehensive study of their performance characteristics and system design trade-offs has been conducted.
This work presents the first systematic analysis of the trade-off between neural network and lookup table size in Safety Nets.

Paper AbstractExpand

The integration of Artificial Intelligence (AI) in safety-critical aviation systems presents significant challenges for certification and deployment. Aviation, often regarded as the safest form of transportation, relies on numerous safety-critical systems. For future safety-critical AI-based systems, EASA requires a Safety-by-Design approach, which can be achieved by using Safety Nets that combine neural network compression with lookup tables to ensure 100 % correct runtime behavior across the discretized operational design domain. Although Safety Nets have been studied, no comprehensive study of their performance characteristics and system design trade-offs has been conducted. This work presents the first systematic analysis of the trade-off between neural network and lookup table size in Safety Nets. By systematically comparing neural networks with diverse architectures, this study identifies optimal design parameters that minimize overall storage and memory requirements while maintaining certification compliance. Results demonstrate that architectures with 3 to 5 hidden layers, each with approximately 50 to 100 nodes, combined with one-hot encoding, achieve the best balance. In these configurations, neural networks accurately represent at least 97 % of the data, while compact lookup tables handle the remaining errors. The resulting Safety Nets reduce the system size by almost three orders of magnitude, fitting within the memory budget of current avionics hardware while guaranteeing 100 % correct outputs across the entire discretized input space, as required by EASA guidelines. This work provides the first-ever open-source implementation of Safety Nets for HCAS and VCAS with replicable results, demonstrating a practical pathway toward certifiable AI-based systems in aviation and establishing Safety Nets as a viable Safety-by-Design solution for safety-critical applications.

This paper addresses the challenge of certifying Artificial Intelligence (AI) for use in safety-critical aviation systems. To meet EASA requirements for a "Safety-by-Design" approach, the authors propose using "Safety Nets"—a hybrid architecture that combines neural networks with lookup tables to guarantee 100% correct runtime behavior within a defined operational domain.

How Safety Nets Work

The Safety Net approach addresses the inherent unpredictability of neural networks by pairing them with a corrective mechanism. The neural network handles the majority of the data processing, while a compact lookup table is used to store and correct any errors the network makes. By discretizing the operational design domain, the system ensures that every possible input is mapped to a verified, correct output. This configuration allows the system to meet strict aviation certification standards while still utilizing the efficiency of neural networks.

Optimal Design Parameters

The researchers conducted a systematic analysis to find the best balance between neural network size and lookup table memory usage. They found that the most efficient configurations utilize:

Neural Network Architecture: 3 to 5 hidden layers, with 50 to 100 nodes per layer.
Encoding: One-hot encoding.
Performance: These configurations allow the neural network to accurately represent at least 97% of the data, leaving only a small fraction to be handled by the lookup table.

Impact on Avionics Hardware

A primary goal of this research was to fit AI systems within the constrained memory budgets of current avionics hardware. The study reports that the Safety Net approach reduces the overall system size by nearly three orders of magnitude compared to standard neural network implementations. This reduction makes it possible to deploy AI-based systems that are both performant and compliant with EASA safety guidelines.

Practical Implementation

The authors provide an open-source implementation of Safety Nets specifically for HCAS (Horizontal Collision Avoidance System) and VCAS (Vertical Collision Avoidance System). By offering replicable results, the work establishes a practical pathway for integrating AI into aviation, moving beyond theoretical models toward certifiable, real-world applications.

Comments (0)

No comments yet

Be the first to share your thoughts!