In safety- and mission-critical embedded systems, particularly autonomous vehicles and aerospace applications, large AI models must deliver high performance while meeting strict power, energy, and reliability requirements under harsh operating conditions. Existing commercial AI accelerators provide strong computational efficiency but limited fault protection, while hardened platforms sacrifice performance and energy efficiency. Likewise, current software-based reliability methods do not scale effectively to large machine learning models.

The Reliable AI for the Next Generation of Safety- and Mission-Critical Embedded Systems (RAISE) project addresses these limitations through a cross-layer approach based on the hypothesis that not all hardware faults affect AI inference equally. RAISE will trace faults across hardware structures and model operations, combine AI-level sensitivity metrics with microarchitectural fault analysis, and use sensitivity-guided selective protection to develop scalable reliability-assessment and fault-mitigation methods. These methods will be validated through microarchitectural simulations and realistic fault-injection campaigns involving radiation and temperature experiments, with failure-rate objectives aligned with terrestrial functional-safety requirements and acceptable rates for space applications, as defined with industrial partners. The core innovation of RAISE is to make AI reliability predictive rather than exhaustive by linking hardware fault locations to AI model failures and using this relationship to guide efficient cross-layer protection.