What question did this study set out to answer?

The aim is to create a machine learning framework that predicts the risk of cracking in additively manufactured nickel-based superalloys.

April 4, 2026Open Access

An interpretable machine learning framework for the prediction of cracking in additively manufactured high gamma-prime nickel-based superalloys

Key Points

The aim is to create a machine learning framework that predicts the risk of cracking in additively manufactured nickel-based superalloys.
Used infrared thermal histories for crack susceptibility analysis in alloy compositions.
Developed a four-module pipeline including thermal history alignment, PCA compression, classification, and probability mapping.
Analyzed 33 deposits across various compositions and process parameters.
Achieved 96% accuracy with Gaussian Process classifier using composition and thermal descriptors.
Low-order principal component scores were most influential in predicting crack susceptibility.
Identified critical susceptibility transitions in the composition-parameter space.

Abstract

Cracking during directed energy deposition limits the processability of components fabricated from high γ′ Ni-based superalloys and demands prediction tools that are both accurate and transparent while computationally efficient. This work develops a lightweight, data-driven framework that correlates in-situ infrared (IR) thermal histories to crack susceptibility in IN625–IN100 graded alloy mixtures spanning compositions with varying γ′-forming precursors (Al, Ti). Thirty-three thin walls were deposited across a range of compositions and process parameters, generating diverse thermal histories for analysis. The pipeline comprises four modules: (i) extraction and alignment of pixel-level IR thermal histories (9 × 27 grid) from identical locations of each deposit, (ii) PCA-based compression of 891 layer-level profiles into low-dimensional thermal-history descriptors (first ten PCs explain > 90% variance), (iii) crack-susceptibility classification using composition/process inputs with and without PC scores (CP-P, CPS-P, and CS-P), and (iv) crack-probability mapping over the composition–parameter space. Using standardized inputs, the best-performing configuration (composition + process + PC scores) with a Gaussian Process classifier (ARDSE kernel) achieved accuracy = 0.96 (F1 = 0.92; recall = 0.92), while composition + PC scores performed comparably, indicating that thermal histories largely subsume nominal process metadata. Feature-relevance analyses (Random-Forest importances and ARDSE per-feature length-scales) consistently ranked low-order PC scores and composition as most influential. The resulting maps reveal sharp susceptibility transitions near ~ 87.5 wt% IN100 under no-dwell and strong mitigation with a 10 s dwell, highlighting the utility of compact thermal descriptors for process planning and in-situ quality assurance in metal additive manufacturing.

Mark Helpful

Bookmark

Relay

View Full Paper