Work

Learning the Ductile-to-Brittle Transition Curve of Steels: Direct Machine Learning Versus Empirical-Equation-Constrained Models with Validation on In-House Cast Steels

Year: 2026
Type: article
Abstract: The Charpy impact energy of ferritic steels varies sigmoidally with temperature, and the resulting ductile-to-brittle transition (DBT) curve is commonly described by a hyperbolic-tangent (tanh) equation. This study compares the following two ways of learning this curve from heterogeneous literature data: direct regression of the impact energy, and an equation-constrained approach in which the learner predicts the four tanh parameters and is trained end-to-end through the equation. A database of 604 Charpy V-notch measurements from 40 publications (58 steels) was compiled and supplemented with physics-derived descriptors, and six tree and neural-network models were tuned under the same protocol. Under random cross-validation the five models that use all measurements reach R2 = 0.71–0.89, but when whole steels are held out, R2 falls to 0.12–0.32. The equation-constrained models return smooth sigmoids with interpretable parameters at a similar accuracy under the grouped schemes, but the transition temperatures predicted for unseen steels remain uncertain (mean absolute errors of 35–41 °C). On four in-house quenched-and-tempered cast steels tested between −40 and 25 °C, the equation-constrained network had the lowest mean absolute error (28 J; other models 44–55 J), a ranking that depends on network settings and on the values assumed for unanalysed elements. Adaptation with three cast steels reduced the error on the fourth to 21–29 J for all models, no lower than a curve fitted to the cast data alone. SHAP attributed the transition temperature and the upper-shelf energy mainly to carbon and tempering.
Source: Metals
Language: English
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7236822661
FWCI:
Cites: 60
Cited by:
Related to: 10