Workshop: Resilient AI at the Space Edge (RAISE-2026)

Workshop Overview

As we enter the era of autonomous lunar outposts and mega-constellations, space missions are transitioning from simple telemetry to autonomous decision-making, the reliance on high-performance AI accelerators at the “edge” (on-orbit, lunar, or deep space) has intensified. The harsh space radiation environment introduces transient and permanent faults—such as Single Event Effects (SEE) and Total Ionizing Dose (TID) degradation—that can subtly corrupt AI training and inference. We are entrusting mission success to AI silicon that is inherently vulnerable. Traditional radiation-hardening is too slow and expensive for the pace of modern AI. RAISE-2026 is the forum for defining how we build ‘Safe-Fail’ autonomous systems using a mix of cutting-edge COTS (Commercial-Off-The-Shelf) hardware and resilient algorithmic frameworks.

Scope and Topics of Interest

We invite presentations, posters, and discussion panels focusing on the detection, mitigation, and testing of hardware-induced faults in AI systems. Key topics include, but are not limited to:

Radiation Hardening by Design (RHBD) for AI accelerators

• Radiation Hardened Design: Design considerations for radiation hardening of GPUs, TPUs, NPUs, and FPGAs.
• Next-Gen AI Hardware: Comparing Rad-Hardened by Design (RHBD) chips vs. Radiation-Tolerant COTS chips in space.
• The “Hidden Fault” Crisis: Strategies for screening manufacturing defects that only manifest under thermal or radiative stress at the edge.
• Neuromorphic and Non-Von Neumann Architectures: Exploring if spiking neural networks are inherently more resilient to bit-flips.
• Reliability Testing, Fault Characterization with Digital Twins
• Virtual Radiation Beams: Methodologies for identifying hidden manufacturing defects vs. radiation-induced faults.
• In-Situ Health Monitoring: In-situ testing protocols for processors in LEO, MEO, GEO orbits, lunar and Mars and deep space.
• Virtual Testbeds: Simulation environments and tools for SW/HW testing. Using digital twins and high-fidelity fault injection to predict hardware failure before launch.
• Cross-Platform Benchmarking: Standardizing how we measure “AI Reliability” across different orbits (LEO vs. Deep Space) under heavy ion and proton beam testing.

Resilient AI Algorithms & Software

• Fault-Tolerant Training: Techniques for training models that are inherently robust to weight corruption.
• Silent Data Corruption (SDC): Detecting and mitigating the impact of bit-flips on deep learning inference.
• Self-Healing Models: Resilient neural networks and dynamic redundancy.
• On-Board Monitoring: On-board telemetry analysis for real-time health monitoring of AI hardware.

Systems & Architecture

• Hardware-in-the-loop (HIL) simulation for space-edge AI.
• Cross-layer resilience: Coordinating software error correction with hardware redundancy.
• Hybrid Inference: Balancing high-power/high-risk inference with low-power/safe-mode logic. Power-efficient fault detection at the edge.
• Case studies: AI anomalies in past or current space missions.

Call for Participation

We encourage contributions from

• Industry: Providers of space-grade SoCs, FPGAs, and AI-optimized hardware.
• Research Organizations: National labs and space agencies (NASA, ESA, JAXA) focusing on mission assurance.
• Academia: Researchers working on the theoretical bounds of AI reliability and computer architecture.

Submission Formats

• Technical Presentation and Panel Discussion: A title and an abstract only, 200 – 500 words.
• Lesson Learned Brief: A 1-2 page summary of a specific failure or testing hurdle.
• Interactive Demo Proposal: Showcasing fault-injection tools or resilient AI frameworks.

Workshop Format:

• One Panel Discussion Session in the morning (1.5 hours each session).
• A poster session & demo in the afternoon.
• Lunch time informal networking.