Physically interpretable crack prediction using dynamic contour point-based frequency-domain gray-level features in laser-directed energy deposition of alumina ceramics

Crack defects remain a critical bottleneck in laser-directed energy deposition (LDED) of ceramics due to the material's intrinsic brittleness, combined with rapid thermal cycling and residual stress accumulation triggered by localized heating. Predicting crack defects is particularly challenging due to the difficulty in real-time measurement of residual stress and the limited stable capture of the transient temperature evolution that governs crack initiation and propagation. To circumvent the need for direct residual stress measurement, this study proposes a machine learning-based crack-rate prediction framework that leverages frequency-domain thermal characteristics of feature points, grounded in the physical mechanism by which cracks propagate when residual stresses exceed critical thresholds. To address the instability introduced by randomly selected fixed-point temperature sampling under the highly dynamic behavior of the molten pool, a novel dynamic contour point (DCP)-based frequency-domain gray-level features extraction method was developed. The DCPs, defined by the evolving deposition contour and corrected with a γ-scaled average thickness of adjacent layers, enable stable and physically representative thermal feature extraction closely correlated with crack formation. Experimental results reveal a pronounced negative correlation between the mean amplitude of the gray-level frequency spectrum and the crack rate. Multiple machine learning models trained on small-sample datasets using DCP-based frequency-domain gray-level features achieved high predictive accuracy, with the Artificial Neural Network reaching 99.48%. This work provides a viable technical pathway for in situ crack-rate prediction and lays the groundwork for adaptive process control in future LDED of ceramics.