
Topology optimization is increasingly important for additive manufacturing because it enables lightweight structures with complex load paths, while process-induced variability can produce spatial heterogeneity in fatigue resistance. This study develops a distributionally robust fatigue-constrained topology optimization framework that distinguishes two sources of uncertainty: spatial variability in local fatigue strength and ambiguity in the probabilities assigned to sampled uncertainty realizations. Spatial variability is represented by a correlated fatigue-strength random field, while probability-model ambiguity is treated using a discrete Kullback–Leibler ambiguity set. The framework is evaluated using an Al 6061-T6 cantilever benchmark and transfer studies involving Ti-6Al-4V, an L-bracket geometry, and a three-dimensional benchmark. In the principal Al case, KL-DRO reduces independent out-of-sample fatigue exceedance from 17.56% to 12.88% and lowers the 95th-percentile fatigue measure from 1.999 to 1.685, with an approximately 2.2% increase in material relative to mean-risk optimization. Similar reductions are observed for the L-bracket and three-dimensional benchmarks, whereas Ti-6Al-4V and matched material studies show little or no advantage. The results therefore indicate that KL-DRO can improve out-of-sample fatigue-tail performance when fatigue risk is active and the empirical uncertainty support is sufficiently informative, but it does not universally outperform mean-risk optimization. The proposed framework provides a computational basis for incorporating spatially resolved fatigue-property uncertainty into topology optimization for additively manufactured structures.
Machine-learning predictions are used to assess material properties, monitor manufacturing processes and support inspection and maintenance. Their engineering value depends on whether the information used by a model can be examined and related to the physical task. This narrative review considers interpretable models, feature attribution, response plots, visual explanations and constrained counterfactuals in manufacturing and materials processing. It distinguishes predictive performance from explanation fidelity, physical evidence and decision usefulness. Selected applications illustrate these differences: process–property modelling in additive manufacturing, acoustic monitoring, tool-condition assessment, ultrasonic defect sizing and uncertainty in production processes. A comparison of these examples shows why the prediction target, data representation and validation procedure must be stated before explanation results are interpreted. The review proposes four categories of explanation use—descriptive, diagnostic, prescriptive and governed—and identifies the additional evidence required for each. The main methodological concerns are dependent inputs, unrealistic perturbations, non-independent observations and changes in the model or operating conditions. Explanation methods can help examine a model and guide further investigation, but process changes require separate checks of feasibility and physical consequences. The narrative selection does not establish the prevalence of individual methods across the field.
Ti-6Al-4V is widely employed for use in aerospace, biomedicine and marine engineering applications. Material removal rate (MRR) in Wire Electrical Discharge Machining (WEDM) is strongly dependent on the wire electrode and discharge settings but has been rarely studied using statistical and Machine Learning (ML) methods together. For this purpose, a Taguchi L27 orthogonal array design was used to study the influence of peak current, pulse-on time, pulse-off time and wire electrode (brass, zinc-coated brass, molybdenum) on the MRR. The very same 27 experimental data points were analyzed using the signal-to-noise ratio method and analysis of variance, as well as using five ML algorithms for regression (polynomial ridge regression, support vector regression, random forest, extreme gradient boosting and artificial neural network) assessed through leave-one-out cross-validation. Peak current, pulse-on time and wire electrode had the strongest influence, whereas pulse-off time was statistically insignificant. The support vector regression approach performed better with an R2 of 0.81 in cross-validation and root mean squared error of 0.31 mm3/min in predicting the MRR. Both optimization methods predicted identical optimal conditions (zinc-coated brass wire, pulse-on time of 111 µs, pulse-off time of 61 µs, and peak current of 220 A). The only exception was an insignificant pulse-off time. Since both analysis methods use the same non-replicated data set (n = 27), the agreement is just a consistency test, not a validation, and the results are valid only for the examined range.
This study investigates the effect of laser remelting on the microstructure, phase composition, mechanical properties, tribological behavior, and corrosion resistance of self-fluxing NiCrFeBSiC coatings deposited onto AISI 321 stainless steel by detonation spraying. The coatings were characterized before and after laser treatment using scanning electron microscopy (SEM), energy-dispersive X-ray spectroscopy (EDS), X-ray diffraction (XRD), instrumented indentation, adhesion testing, ball-on-disk tribometry, and potentiodynamic polarization measurements. SEM analysis revealed that laser remelting reduced coating porosity from 0.306 ± 0.014% to 0.15 ± 0.09%, accompanied by microstructural densification and a reduction in coating thickness. XRD analysis confirmed the preservation of the principal crystalline phases, including Ni3Fe, Cr7C3, CrB, and Ni3B, following laser treatment. The Vickers hardness increased from 1049.1 ± 43.4 to 1206.1 ± 63.1 HV, while the elastic modulus increased from 233.6 ± 6.2 to 258.4 ± 5.3 GPa. Furthermore, adhesion strength increased by 19.4%, whereas surface roughness decreased by 19.2%. Laser remelting significantly improved tribological performance, reducing the average coefficient of friction from 0.535 ± 0.048 to 0.409 ± 0.086 and the specific wear rate from 2.864 × 10−4 to 1.447 × 10−4 mm3/(N·m), corresponding to a 49.5% reduction in wear. Electrochemical measurements indicated a modest decrease in corrosion current density of approximately 3.0%, accompanied by changes in the morphology of corrosion-affected regions. The improvements are associated with laser-induced microstructural densification, redistribution of alloying elements, and preservation of hard reinforcing phases. These findings demonstrate the potential of combining detonation spraying with laser remelting to produce wear-resistant nickel-based protective coatings for mechanical and power engineering applications.
Seamless copper tube manufacturing involves sequential processing steps, while the final product quality is assessed through Eddy Current Testing (ECT). Predicting this quality early on using billet heating and hot-extrusion data poses a challenge as the underlying models might exploit production-related signatures instead of transferable process relationships. This study investigates if the production data describing these two stages can predict the classified ECT quality by utilizing random-forest models against different validation strategies, aiming to assess the historical performance, the actual information learned and if this information transfers across tools and time. The analysis covers 1916 observations from 5 months, three products, 62 mandrels and 155 dies at a single plant. Three ECT-based quality classes (A, B, C) were analyzed. Random cross-validation yielded balanced accuracies of 47.49% with all predictors and 46.01% with physical predictors (permutation p = 0.0033 for both). Mandrel-grouped scores were 39.90% and 41.31%, respectively, while the all-predictor score decreased from 50.12% under random holdout to 35.00% on future production. Physical variables retained historical signal, but a stable tool-independent relationship was not established. These findings provide an initial basis for developing context-aware quality-risk and model-applicability assessment, while retaining ECT as the final inspection.