Deep Learning Unlocks X-ray Microtomography Segmentation of Multiclass Microdamage in Heterogeneous Materials.

Area of Science:

• Materials Science
• Computational Science
• Engineering

Background:

• Four-dimensional quantitative characterization of heterogeneous materials using in situ synchrotron radiation computed tomography (SRCT) reveals 3D sub-micrometer features, especially evolving damage under load.
• Increasing dataset size and complexity necessitate time-intensive and subjective semi-automatic segmentations for accurate analysis.

Purpose of the Study:

• To present the first deep learning (DL) convolutional neural network (CNN) segmentation for multiclass microscale damage in heterogeneous bulk materials.
• To develop an objective and efficient method for analyzing complex composite damage, overcoming limitations of traditional segmentation techniques.

Main Methods:

• Training a CNN on approximately 65,000 human-segmented tomograms of advanced aerospace-grade composite damage.
• Applying the trained CNN for segmentation of complex and sparse microscale damage features within tomographic datasets.

Main Results:

• The CNN achieved ≈99.99% agreement in segmenting composite damage classes, significantly outperforming traditional algorithms.
• Nearly 100% of human segmentation time was eliminated, with the machine often outperforming human analysis by discovering new damage and improving segmentation in artifact-rich areas.
• DL proved to be a disruptive approach, accelerating knowledge creation by two orders of magnitude through generalizable, ultrahigh-resolution feature segmentation.

Conclusions:

• DL-based segmentation offers unprecedented objectivity and efficiency for quantitative structure-property characterization of materials.
• This approach enables high-throughput knowledge creation by automating and enhancing the analysis of microscale damage in complex materials.
• The generalizability of the DL model suggests broad applicability in advanced materials characterization and development.