Citation

Tynchenko, Valeriya V. and Kurashkin, Sergei O. and Bukhtoyarov, Vladimir V. and Connie, Tee and Borodulin, Aleksei S. and Mikhailov, Artem. Y. and Tynchenko, Vadim S. (2026) Multi-output machine-learning benchmark and Pareto-based decision support for electron-beam welding of thin-walled titanium structures with comparative synthetic-data augmentation. The International Journal of Advanced Manufacturing Technology. ISSN 0268-3768

Text s00170-026-18892-y.pdf - Published Version Restricted to Repository staff only Download (4MB)

Abstract

Electron-beam welding (EBW) of thin-walled titanium components is a key joining technology for aerospace and highvacuum applications, in which reliable data-driven prediction of weld-bead geometry from controllable process parameters underpins shop-floor decision support. Existing studies are nevertheless constrained by small experimental data sets, narrow algorithmic coverage, ad-hoc hyperparameter tuning, scalar evaluation criteria and the absence of a formal statistical comparison framework. The present work proposes a unified methodology that addresses these gaps through (i) a comparative evaluation of four synthetic-data generators—conditional tabular GAN, tabular variational autoencoder, Gaussian copula, and a physics-informed generator based on the moving-source thermal solution—with a convergence study from 1 000 to 50 000 samples and a five-mode ablation protocol (real-only, synth-only, real+synth, train-on-synthetic–test-onreal, train-on-real–test-on-real); (ii) a 41-model multi-output regression benchmark spanning generalised linear, kernel, ensemble, gradient-boosting and tabular deep-learning families; (iii) a 12-method hyperparameter optimisation comparison that includes three genetic-algorithm variants and the multi-objective NSGA-II; (iv) a 15-metric evaluation panel reported with bootstrap confidence intervals; (v) Friedman–Nemenyi and Wilcoxon signed-rank statistical comparison; (vi) Paretobased inverse parameter design for joint depth-and-width targeting; and (vii) global and local interpretability through SHAP, partial-dependence and accumulated-local-effects analyses. On the real ablation a Friedman test rejects equality of the 41 regressors (χ2 = 439.0, p =1.3 × 10−68); the ten best models lie within 0.013 macro-R2 of one another (NGBoost leading at 0.936), and only NGBoost and the mixture density network are Pareto-optimal in the joint (RMSED, RMSEW ) sense. Among synthetic generators, TVAE, the Gaussian copula and the physics-informed generator all yield usable downstream models (median macro-R2 of 0.85, 0.74 and 0.81), whereas CTGAN trained on synthetic data alone falls below a constant predictor at N = 72. The NSGA-II inverse design recovers admissible process windows for prescribed depthand-width targets within the observed envelope, and additionally maps the achievable depth–width capability frontier. The framework is released as open-source code with a Zenodo-archived release, supporting reproducibility and providing a methodological template for other small-data manufacturing-process problems

Item Type:	Article
Uncontrolled Keywords:	Synthetic-data augmentation, Pareto-based decision support, Machine learning
Subjects:	Q Science > QA Mathematics > QA71-90 Instruments and machines > QA75.5-76.95 Electronic computers. Computer science
Divisions:	Faculty of Information Science and Technology (FIST)
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	04 Sep 2026 06:37
Last Modified:	04 Sep 2026 06:37
URII:	http://shdl.mmu.edu.my/id/eprint/16724

Downloads

Edit (login required)