The research team assembled a database of 396 tensile testing records for sub-sized Stainless Steel 316 specimens, drawn from an extensive review of peer-reviewed literature and scientific databases. Each record was described by 45 input features spanning material composition, manufacturing and heat treatment history, irradiation conditions, specimen geometry, and testing parameters such as temperature and strain rate. The team set out to predict four key tensile properties: ultimate tensile strength, yield strength, total elongation, and uniform elongation. After cleaning to remove records with missing values, the usable datasets shrank further—to just 209 points for ultimate tensile strength, 204 for yield strength, 190 for total elongation, and a mere 86 for uniform elongation. For a field where safety margins depend on reliable property predictions, those numbers are dauntingly small.

The imbalance in the data was as troubling as its sparsity. Yield strength values, for example, showed a distinctly bimodal distribution with a conspicuous gap between roughly 450 and 500 megapascals. The low-strength cluster corresponded to solution-annealed specimens treated above 1000 degrees Celsius, while the high-strength cluster came from cold-worked and stress-relieved material. This kind of clustering means a model can be trained heavily on one regime and fail badly in another, particularly in the underrepresented regions of the feature space that matter most for understanding how specimen size affects measured properties. Size effects themselves complicate matters: when specimen dimensions approach microstructural length scales such as grain size or precipitate spacing, deformation mechanisms shift from bulk-dominated to surface-mediated behavior, and conventional models of deformation and fracture can lose validity.

To attack the data problem, the researchers turned to two very different augmentation strategies. The first was a Generative Adversarial Network, or GAN, in which a generator network produces synthetic feature vectors from random noise while a discriminator network learns to distinguish real from fake samples, forcing the generator to create increasingly realistic data. The team trained the GAN for 10,000 epochs using the Adam optimizer, with fully connected layers, batch normalization, and LeakyReLU activations. The second method, SMOGN—the Synthetic Minority Oversampling Technique for Regression with Gaussian Noise—works more directly: it interpolates between existing samples and their nearest neighbors, then perturbs the results with Gaussian noise, targeting the sparsest regions of the target variable’s distribution. Synthetic compositions were constrained to physically realistic ranges, with elemental fractions normalized to sum to 100 percent. Both methods were applied at augmentation ratios of 20, 40, 60, 80, and 100 percent additional synthetic data.

Nine machine learning algorithms were then evaluated in a remarkably thorough experimental design: 396 individual model configurations, each assessed with five-fold cross-validation using the Pearson correlation coefficient, the coefficient of determination, and root-mean-square error. The algorithms spanned the modern machine learning spectrum, from Random Forest and Extreme Gradient Boosting (XGBoost) to decision trees, k-nearest neighbors, support vector machines, Gaussian process regression, neural networks, and two transformer architectures—one treating each feature as a token, and a Materials-Informed Transformer that grouped features by physics-based categories such as composition, processing, and irradiation conditions.

The results delivered a clear verdict: tree-based ensemble methods dominated. Random Forest achieved a Pearson correlation of 0.987 for yield strength with a root-mean-square error of 32.6 megapascals, while XGBoost reached 0.989 for uniform elongation with an error of just 3.414 percent—extraordinary accuracy given that the uniform elongation dataset contained only 86 points. The researchers took care to verify that this performance reflected genuine learning rather than statistical luck: a naive baseline that simply predicts the training-set mean produced a root-mean-square error of 21.714 percent and a negative coefficient of determination, and permutation importance analysis showed that the most influential features were heat treatment condition and test temperature, both of which have well-established physical relevance to ductility. Support vector machines fared worst, with yield strength correlation dropping to 0.892 and errors 27 percent higher than Random Forest, underscoring how poorly linear decision boundaries capture the coupled, non-linear relationships in materials data.

Perhaps the most intriguing finding was that data augmentation is strongly property-dependent. For strength-related properties, which are governed by discrete and competing mechanisms such as solid-solution strengthening, precipitation, and cold work, GAN-generated samples consistently helped. XGBoost improved by about 1 percent for ultimate tensile strength at 60 percent GAN augmentation, and k-nearest neighbors—normally expected to be hypersensitive to synthetic data—gained 1.4 percent at full GAN augmentation, apparently because the generator produced chemically realistic intermediate compositions that improved local neighborhood quality. For elongation properties, which follow smoother, more continuous geometric scaling trends, SMOGN’s interpolative approach worked better. XGBoost peaked at 80 percent SMOGN augmentation for total elongation, and a decision tree jumped 1.7 percent for uniform elongation at 60 percent SMOGN. Meanwhile, some combinations backfired badly: SMOGN degraded Random Forest’s ultimate tensile strength predictions by over 8 percent, and support vector machines deteriorated under both augmentation types across the board.

The study also delivered a cautionary tale for the physics-informed machine learning community. The team built Physics-Informed Neural Networks embedding two classical empirical relationships—Barba’s Law and the Bertella-Oliver formula—as constraints on total elongation prediction. Intuitively, adding physics should help a data-starved model. Instead, both PINN variants underperformed a standard neural network by roughly 10 percent in correlation, with root-mean-square errors rising from 9.87 percent to 15.4 and 16.0 percent. The authors attribute this to a mismatch between the rigid, globally applied empirical laws and a heterogeneous dataset spanning multiple alloys and processing conditions. When the imposed physics is wrong for parts of the data, it actively hurts. The lesson is that physics constraints must be flexible and data-adaptive—learning material-dependent parameters or using soft constraints—rather than hard-coded relationships.

Transformer architectures told a more nuanced story. The Simple Tabular Transformer ranked third for yield strength at 0.979 correlation but fell to eighth place for ultimate tensile strength, a 14 percent performance gap revealing its sensitivity to underlying data structure. The Materials-Informed Transformer, with its physics-based feature grouping, consistently underperformed its simpler cousin by 2 to 15 percent, and ablation experiments ruled out grouping granularity as the culprit—suggesting the prescribed attention structure conflicted with data-driven patterns. Interestingly, the MIT showed the largest relative gains from augmentation, improving 4.4 percent for total elongation under SMOGN, hinting that physics-based architectures may simply need more training diversity to learn useful attention patterns.

The practical implications extend well beyond Stainless Steel 316. Nuclear material qualification traditionally demands large specimens and extensive irradiation campaigns that can stretch over years, and miniaturized testing is one of the most promising routes to accelerating the screening of candidate alloys for advanced reactors. By demonstrating that carefully chosen algorithm-augmentation pairings can extract reliable predictions from a few hundred records, the study offers a template for data-starved materials science more broadly: match the augmentation strategy to the physics of the property being predicted, favor ensemble methods for stability, and treat rigid physics constraints with skepticism when the data spans diverse regimes. The team’s framework, validated through cross-validation stability analysis showing XGBoost’s standard deviation as low as 0.003 for uniform elongation, points toward faster, cheaper, and safer qualification of the materials that next-generation nuclear reactors will depend on.

Subject of Research: Machine learning and data augmentation for predicting tensile properties of sub-sized nuclear structural materials

Article Title: Predicting tensile properties of sub-sized nuclear materials using data augmentation strategies on imbalanced and sparse datasets

Article References: Smagulova, A., Vakanski, A., Merickel, J. W., Lee, I., Purushotham, C., Tang, Y., Cerpa, A. E., Xu, F., & Song, R. (2026). Predicting tensile properties of sub-sized nuclear materials using data augmentation strategies on imbalanced and sparse datasets. Frontiers in Manufacturing Technology, 6, Article 1811951. https://doi.org/10.3389/fmtec.2026.1811951

Image Credits: AI Generated

DOI: 10.3389/fmtec.2026.1811951

Keywords: machine learning, data augmentation, nuclear materials, tensile properties, Stainless Steel 316, GAN, SMOGN, Random Forest, XGBoost, sub-sized specimens, physics-informed neural networks, materials informatics