NL-MambaXCT hits 96.9% accuracy for aerospace defect detection

Industrial non-destructive testing of Nomex honeycomb structures — widely used in aerospace manufacturing — still relies heavily on manual inspection or supervised models that require large labeled datasets. To address this, Ghaleb Aldoboni and colleagues introduced NL-MambaXCT, a framework that integrates a Mamba-based backbone with self-supervised masked image modeling (MIM) and a novel Nested Learning (NL) formulation. The backbone uses a four-stage 2D encoder: early stages employ RegNet convolutional blocks, while deeper stages incorporate Mamba-based sequence mixing with attention. The model is first pre-trained on 19,961 unlabeled industrial XCT slices using MIM, then fine-tuned on only 2,000 relabeled Nomex XCT slices split by production order, drastically reducing the need for costly manual annotations.

NL-MambaXCT's key innovation is nested learning, implemented via two-timescale parameter dynamics. Selected projections maintain slow exponential-moving-average traces alongside fast weights, while a deep-momentum optimizer adds a second slow parameter-update trajectory — enabling the model to capture both fine-grained details and long-range dependencies. On a held-out test set, the MIM-pretrained model achieves 96.91% accuracy and 96.8% macro F1, outperforming CNN, attention, and single-timescale Mamba baselines by 3.11 to 10.31 percentage points in accuracy. This work demonstrates that combining masked self-supervision with nested learning dynamics is a promising strategy for robust defect classification in Nomex honeycomb XCT inspection, particularly where labeled data is scarce.