.png){kind=icon}
.png){kind=icon}
Return
Knowledge-guided causal analog learning for small-sample steel property prediction with process-response reasoning
.png){kind=icon}
PRE
AI
C
D
W
Z
Y
Z
J
G
DOI：10.1016/j.mtla.2026.102809.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Small experimental steel datasets challenge property prediction because mechanical properties depend on coupled composition-processing-microstructure relationships and often require extrapolation to unseen alloys. We propose a knowledge-guided causal analog learning framework for small-sample steel property prediction with counterfactual process-response reasoning. Metallurgical knowledge is implemented as a lightweight, fixed descriptor graph that maps alloying elements and heat-treatment variables to hardenability, tempering-response and composition-process interaction descriptors; these descriptors define both the predictive feature space and the analog similarity space. On the cleaned medium-carbon steel dataset, which contained 116 sample and 32 composition groups, the full model achieved an average leave-one-composition-out (LOCO) R2 of 0.655 across yield strength, ultimate tensile strength, elongation, reduction of area and impact energy. Knowledge-guided target-coupled heads were the dominant source of improvement, while analog weighting provided a smaller but useful refinement over the no-analog variant. Within-composition process-response tests gave direction accuracies of approximately 0.80, whereas a process-property permutation placebo was close to chance (0.497). External validation on a public low-alloy steel composition-temperature-property dataset (913 cleaned rows and 95 alloys) further showed that knowledge-guided descriptors improved LOCO performance over composition-only and composition-temperature baselines.
Journal
IF:
2.9
Papers:
2.2K
Citations:
6.5K
Organization
No organization information available
Cited Papers