A Radiation-Hardened, Fault-Aware Quantized Neural Network That Self-Heals

Peter Xiangyuan Ma, Joshua S. BloomarXiv 2026 · arXiv

TL;DR

This paper introduces Projection Healing, a software-based self-repair mechanism that allows quantized neural networks to recover from radiation-induced bit flips on standard, off-the-shelf hardware. By training models to concentrate their internal representations into a low-rank subspace, the system can detect corruption at runtime and reconstruct lost data from uncorrupted "survivors," doubling the operational lifespan of AI models in high-radiation environments like space.

What Problem Does This Paper Address?

Deep learning models are increasingly deployed in "radiation-challenging" environments—think satellites in low-Earth orbit, particle accelerators, or nuclear facilities. In these settings, high-energy particles can ionize silicon, causing bits in memory to flip.

While hardware designers often use Triple Modular Redundancy (TMR) or Error-Correcting Code (ECC) memory to combat this, these solutions are expensive, power-hungry, and often unavailable on the cheap, commercial off-the-shelf (COTS) hardware used in small satellites. Furthermore, as transistors shrink, the risk of Multi-Bit Upsets (MBUs)—where a single particle strike flips a cluster of adjacent bits—increases. Standard ECC (like SECDED) often fails against these multi-bit bursts. The authors aim to provide a software-only, architecture-independent way to protect quantized neural networks (QNNs) against these catastrophic failures.

Methods

The authors propose a two-pronged approach: Fault-Aware Training (FAT) with a new twist, and a runtime Projection Healing mechanism.

  1. Alignment Penalty (Training): During training, the authors force the network to route 95% of its signal into a pre-defined, low-rank "Green space" using a Discrete Cosine Transform (DCT-II) basis. This creates a mathematical "Red space" where radiation-induced noise is likely to land, making it easier to distinguish signal from corruption.
  2. Projection Healing (Inference): At runtime, the system performs a two-pass check on weight rows and activations.
  3. Pass 1: It calculates a "syndrome score" to detect if a vector has drifted too far into the "Red space."
    • Pass 2: If corruption is detected, it surgically down-weights the flagged coordinates and reconstructs them using the surviving data in the low-rank subspace.
Figure 1. System architecture. (a) Two-pass projection healing on a row. The first DCT-II reconstruction forms the syndrome, the gate sets the blend weights, and a second projection of the surviving coordinates supplies the repaired entries. The weight-alignment term shown is Ew red = max(0, ¯rw −ρ). (b) Training uses four forwards: a clean pass, a contiguous input-block mask, a contiguous weight-block mask, and a contiguous weight sign flip. The range penalty is analytic on the weight rows, and the loss keeps the projection-healing terms separate from the fault-aware terms. (c) A hidden layer heals its weight rows, heals the pre-activation, and then applies the nonlinearity. The drawing continues through the next weight matrix to the clean output.

This method is "memory-light" because it doesn't require storing extra copies of the weights; it uses the inherent redundancy created by the low-rank alignment.

Datasets

The authors evaluated their method on six regression-specific benchmarks, focusing on tasks where maintaining precise output magnitudes is critical:

  • California Housing: Predicting house values.
  • Friedman #1: A synthetic nonlinear regression target.
  • SDSS Photometric Redshift: Predicting astronomical redshift from galaxy colors.
  • NASA C-MAPSS FD001: Predicting the remaining useful life of turbofan engines.
  • NRC Critical Heat Flux: Predicting boiling-crisis thresholds in nuclear reactors.
  • CMS Jet $p_T$: Regressing jet momentum for high-energy physics experiments.

Key Insights and Ideas

  • The "Alignment Tax": Robustness isn't free. Forcing a model into a low-rank subspace incurs a small performance penalty on clean data (a median of +0.3% RMSE). However, this "tax" is a bargain compared to the massive error spikes seen in unprotected models under radiation.
  • MBUs vs. SBUs: While classical SECDED is excellent for Single-Bit Upsets (SBUs), it collapses quickly when MBUs become dominant. Projection Healing is specifically designed to handle these multi-bit bursts by treating the entire vector holistically.
  • Sign Bit Sensitivity: The authors confirmed that the sign bit (bit 7 in INT8) is the most sensitive to corruption. However, they found that protecting only the sign bit isn't enough; the cumulative effect of flips in lower-order bits still leads to model failure at high radiation intensities.
Per-bit position curves showing that while the sign bit (bit 7) dominates damage, lower-order bits still contribute significantly to total error.

Key Results

  • Extended Scrub Intervals: The method extends the time a model can operate before needing a "golden-copy" reload by a median of 2.04× at a 5% error threshold, significantly outperforming SECDED (1.62×).
  • Error Reduction: Toggling the healing mechanism on a fixed checkpoint resulted in a ~23% reduction in the excess error integral across all datasets.
  • Resilience to Intensity: As shown in Figure 2, "Our Model" (red diamonds) consistently stays below the error curves of other software defenses (like Clipped or Ranger) and classical ECC as radiation intensity increases.
MBU sweeps across six datasets showing the proposed model maintaining lower RMSE for longer compared to baselines.

Limitations and Future Work

  • Training Requirement: Projection Healing is not a "plug-and-play" solution for existing models. It requires training from scratch (or significant fine-tuning) to enforce the low-rank alignment.
  • Computational Overhead: The healing process adds arithmetic complexity. At a batch size of 1, the authors report a 2.5× increase in FLOPs compared to a standard unprotected layer.
  • Dataset Sensitivity: One dataset (Critical Heat Flux) showed a much higher "alignment tax" (+11.8% RMSE), suggesting that some tasks are less amenable to low-rank compression than others.
  • Future Directions: The authors suggest exploring different bases beyond DCT-II and investigating how this method scales to much larger models like Transformers, where memory footprints are even more vulnerable.

Summary updated

Abstract

Deep learning (DL) models are increasingly being deployed into radiation-challenging environments in spacecraft, particle accelerators, and, soon, space-based data centers. Instantiated as a learned orchestration of numbers stored in persistent memory, models are susceptible to continual degradation as high-energy particles corrupt data. At inference time, new data passed through the network fundamentally break the off-line assumptions and guarantees of a static model determined during training. To preserve the quality and efficacy of the models, fault tolerance against bit flips is a clear objective. In this work we developed a runtime method for repairing the quantized weights of neural networks (NNs) after bit flips on commodity, commercial off-the-shelf (COTS) hardware, without storing extra replicas. Our work extends the well-established Fault-Aware Training (FAT) paradigm by adding an alignment penalty to concentrate undamaged activations and weights into a low-rank subspace, and we then reconstruct flagged corrupted coordinates from the uncorrupted survivors. We call this method {\it Projection Healing} and apply it to six regression benchmarks on quantized NNs. Holding a FAT checkpoint fixed and toggling only the runtime heal gives a further ${\sim}23\%$ reduction in the native-unit excess integral. Across those datasets the method extends the operational scrub interval by a median of $2.04\times$ at a $+5\%$ RMSE degradation and $1.96\times$ at $+50\%$, compared with $1.62\times$ and $1.07\times$ for classical Single Error Correction, Double Error Detection (SECDED). This introduces a new path towards designing resilient data-driven models, complementing established hardware redundancy, ECC, and FAT-based techniques.
BibTeX
@misc{ma2026radiation,
  author = {Peter Xiangyuan Ma and Joshua S. Bloom},
  title = {{A Radiation-Hardened, Fault-Aware Quantized Neural Network That Self-Heals}},
  year = {2026},
  eprint = {2610.27621},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.27621},
}
Figure