Machine learning reveals hidden pitfalls in predicting residual stress from tiny datasets

Laser cladding has quietly become one of the most consequential technologies in modern manufacturing. By melting powder onto a substrate with a focused laser beam, engineers can repair aerospace turbine components, restore worn energy-sector parts, and fabricate functional coatings for medical implants, all while achieving metallurgically bonded layers with low dilution and enhanced wear and corrosion resistance. Yet the very physics that makes the process powerful also makes it dangerous. The intense, highly localized thermal cycles generate steep temperature gradients and non-uniform cooling rates, which inevitably imprint complex residual stress fields into the clad layer and the underlying substrate. When those stresses are tensile and excessive, they promote crack initiation, distortion, and premature failure, critically undermining the structural integrity and service lifetime of laser-clad components.

For decades, the industry has relied on two main routes to assess these hidden stresses. Experimental techniques such as blind-hole drilling, the contour method, and X-ray diffraction deliver direct measurements, but they are time-consuming, costly, and often limited by surface conditions or geometric complexity. Finite element simulations, by contrast, illuminate the thermo-mechanical mechanisms governing stress evolution, yet they are computationally intensive, highly sensitive to constitutive models and boundary conditions, and demand careful calibration against experimental validation. Machine learning has emerged as a compelling third route, promising rapid nonlinear mappings between process parameters and material responses. But a new study published in Results in Engineering argues that the field has been measuring its own success with a dangerously optimistic ruler.

A research team led by Youbin Lai, Li Zhou, Yujie Li, Jiajun Wu, Xuejun Liu, and Wangwang Ding constructed a dataset of 57 residual stress records extracted from seven independent literature sources on laser cladding of 316L stainless steel systems. The records spanned experimental measurements, experimentally validated simulations, and pure numerical simulation data, capturing laser power, scanning speed, powder feed rate, deposited layer number, and material type as input variables. Rather than chasing maximum predictive accuracy, the team built a robustness-oriented framework that systematically evaluates generalization, stability, and interpretability under the data-scarce conditions that dominate real engineering practice.

The methodological centerpiece is a source-grouped cross-validation strategy. Because the records originate from different studies with heterogeneous material systems, processing conditions, and measurement conventions, conventional random cross-validation can unintentionally place samples from the same literature source into both training and testing subsets. This creates source-related information leakage, allowing models to memorize study-specific patterns rather than learn universal physics. The researchers therefore partitioned data by literature source using GroupKFold and Leave-One-Group-Out validation, ensuring that every held-out source remained completely unseen during training. They also imposed physical consistency criteria, restricting the primary Harmonized dataset to 35 records from four sources where residual stress was defined as the σx component, evaluated at the cladding surface, and obtained from experimental blind-hole measurements.

The results deliver a sobering reality check. Under random 10-fold cross-validation, linear regression achieved a pooled out-of-fold R² of 0.852 and random forest reached 0.715, suggesting respectable predictive power. But when the same models were evaluated with 4-fold GroupKFold by literature source, performance collapsed catastrophically. Linear regression’s pooled out-of-fold RMSE surged from 14.128 to 81.790 MPa, a 478.9 percent increase, while random forest’s error jumped from 19.616 to 64.830 MPa, a 230.5 percent degradation. Both models fell below the Dummy baseline, a non-learning estimator that simply predicts the mean residual stress of the training set. Only support vector regression with an RBF kernel retained a positive pooled out-of-fold R² of 0.147, showing the smallest degradation across validation strategies and the most concentrated fold-level error distributions.

The study also performed source-held-out SHAP analysis to assess whether feature importance rankings remain stable across heterogeneous data sources. Across the four held-out sources, scanning speed and layer number emerged as the most consistently influential features, with scanning speed ranked first in 75 percent of held-out sources and appearing within the top three in all four. Layer number was ranked first in 25 percent of sources and remained within the top three in all. The impurity-based feature importance of the final random forest model, refitted on the complete Harmonized dataset, independently confirmed this hierarchy, assigning 83.4 percent of total importance to layer number and scanning speed combined. However, the magnitude of feature attribution varied substantially across sources, particularly for layer number, and no clear correspondence was observed between attribution magnitude and source-level prediction error.

The researchers also compared three dataset configurations to probe whether data curation alone could resolve the generalization problem. The Full dataset contained all 57 samples from seven sources, the Extended dataset held 43 samples from five sources after excluding two pure simulation studies, and the Harmonized dataset restricted analysis to 35 physically consistent experimental records. Across all three schemes, linear regression and random forest consistently produced negative pooled out-of-fold R² values under source-grouped validation, while support vector regression maintained positive values in both the Extended and Harmonized configurations. Excluding simulation-derived data did not restore positive performance for the two models that fell below the baseline, indicating that literature-derived datasets cannot be assumed to be independently and identically distributed.

The engineering implications extend well beyond model selection. The study argues that standardized data reporting, traceable data provenance, and explicit source metadata are essential for meaningful integration of results across studies. Model evaluation should report pooled out-of-fold performance, fold-level variability, and baseline comparisons alongside conventional accuracy metrics, and sensitivity analyses based on different data curation criteria should clarify whether observed behavior is robust to changes in dataset composition. Predictions from heterogeneous literature datasets are more appropriately used to support parameter screening, trend identification, or hypothesis generation than to replace targeted experimental validation under the relevant material, process, and geometric conditions.

What makes this work resonate far beyond laser cladding is its methodological honesty. The framework does not claim to identify a universally optimal prediction model. Instead, it demonstrates that validation strategy, not model complexity, decides whether machine learning predictions can be trusted when data is scarce and heterogeneous. As data-driven modeling spreads across materials science, additive manufacturing, and engineering at large, the study offers a viral lesson: apparent predictive power can evaporate when the evaluation protocol reflects the deployment setting rather than a favorable data partition. For a field racing to embed artificial intelligence into safety-critical manufacturing, that lesson could not arrive at a more important moment.

Subject of Research: Robust machine learning prediction of residual stress in laser cladding under small-sample conditions

Article Title: Robust machine learning prediction of residual stress in laser cladding under small-sample conditions

Article References: Lai, Y., Zhou, L., Li, Y., Wu, J., Liu, X., & Ding, W. (2026). Robust machine learning prediction of residual stress in laser cladding under small-sample conditions. Results in Engineering, 32, Article 113597. https://doi.org/10.1016/j.rineng.2026.113597

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113597

Keywords: Robust, machine, learning, prediction, residual, stress, laser, cladding, under, small-sample, conditions, scientific research