Imagine trying to get a driver's license for a computer that can't explain why it turned left. Most AI is a 'black box,' which is too risky for trains or satellites. This work creates a way to make AI predictable and traceable, like a flight recorder that proves the system is following safety rules. It ensures the AI knows when it is guessing and can be trusted by regulators.
AI is often a 'black box,' making it impossible to certify for safety-critical use in cars, trains, and satellites. This prevents companies from deploying advanced autonomous features due to strict functional safety regulations.
A safety lifecycle for AI training, a middleware for high-performance platforms, and a set of safety patterns to detect AI anomalies and quantify uncertainty.
If you are an autonomous vehicle manufacturer dealing with strict safety certification for self-driving features — this project developed a safety lifecycle and architecture that allows DL software to meet functional safety standards. This reduces the risk of certification failure for AI-driven road systems.
If you are a satellite operator dealing with the risk of unpredictable AI behavior in orbit — this project developed tools to quantify uncertainty and detect anomalies in AI inference. This ensures space-based AI can be qualified and certified for critical missions.
If you are a train control system provider dealing with the need for deterministic software in autonomous signaling — this project developed a specific middleware and safety patterns for different criticality levels. This enables the adoption of AI in railway systems without compromising passenger safety.
Based on available project data, no specific commercial pricing or licensing costs are mentioned; the project was funded with a EUR 3,891,875 EU contribution.
Yes, the project used industrially-relevant platforms and applied its solutions to three domain-specific case studies in automotive, space, and railway sectors.
The project realized a generic 'core demo' that is fully open source to ease adoption, though specific IP for industrial partners is not detailed.
It provides a way to align Deep Learning software with Functional Safety (FUSA) requirements, making it possible to qualify and certify AI products under existing standards.
The project developed a specific middleware and methods to use high-performance platforms in a predictable and traceable manner.
The consortium is highly balanced for commercialization, featuring a 50% industry ratio with 3 SMEs and 3 research centers. This mix ensures that the technical expertise from centers like BSC and RISE is directly applied to real-world needs of automotive, space, and railway partners, reducing the gap between lab research and industrial deployment.
Contact the Barcelona Supercomputing Center (BSC) regarding the SAFEXPLAIN project outcomes.
Contact SciTransfer to connect with the SAFEXPLAIN consortium for licensing the open-source core demo.
