Machine and Deep Learning in Mechanics of Materials: The Evaluation Protocol Governs Reported Accuracy in Alloy Yield Strength Prediction
6th International Conference on Modern and Advanced Research, Konya, Türkiye, 16 - 17 Ağustos 2026, ss.166-172, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Basıldığı Şehir: Konya
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.166-172
- Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
- Recep Tayyip Erdoğan Üniversitesi Adresli: Evet
Özet
Machine learning is now widely applied to structure and property problems in the mechanics of materials. The accuracy that a study reports, however, depends on how the test set is formed. This work quantifies that dependence for the yield strength of multi-principal element alloys. An open access dataset of 1545 experimental records was used, of which 1067 report both yield strength and test temperature. Eleven composition descriptors were derived from solid solution theory, including atomic size mismatch, mixing entropy and valence electron concentration. Six statistical and machine learning regressors and one deep neural network were trained under three cross-validation protocols. The first protocol shuffles records at random, which is the common practice. The second and the third protocols keep the folds disjoint in alloy composition and in source publication, respectively. A variance component analysis showed that alloy identity explains 64.4 per cent of the yield strength variance, while publication identity explains 61.0 per cent. Random splitting therefore places replicate records of one alloy in the training and the test fold at the same time. Gradient boosting reached a coefficient of determination of 0.796 under random splitting, 0.701 under composition-disjoint splitting and 0.581 under source-disjoint splitting. The deep network followed the same pattern, and the ranking of the models also changed with the protocol. Headline accuracies obtained by random splitting therefore overstate how well a model generalises to an unseen alloy. Grouped validation is recommended whenever a data driven model is intended to support alloy design.