Computational modeling and interpretability analysis of manufacturing parameters in titanium matrix composites

TY - JOUR

T1 - Computational modeling and interpretability analysis of manufacturing parameters in titanium matrix composites

AU - Wei, Qichao

AU - Zhang, Hongmei

AU - Cheng, Xingwang

AU - Zhao, Pingluo

AU - Fan, Qunbo

AU - Mu, Xiaonan

AU - Wang, Yu

AU - Wang, Huaikun

AU - Zhang, Jiaqi

N1 - Publisher Copyright: Copyright © 2026. Published by Elsevier B.V.

PY - 2026/4/5

Y1 - 2026/4/5

N2 - The design of high-performance titanium matrix composites (TMCs) often faces the challenge of complex, nonlinear interactions between multi-scale processing parameters and mechanical properties. To overcome the limitations of traditional empirical methods, this study establishes a comprehensive data-driven computational framework to predict and optimize the performance of TMCs. To address the challenge of small-sample generalization, a diversity-oriented independent validation strategy was employed, strictly separating model training from a heterogeneous hold-out test set. Among various evaluated algorithms, the gradient boosting (GB) model demonstrated superior fidelity, achieving high predictive accuracy on the unseen independent datasets (R2 up to 0.88 for strength). To ensure model transparency, a synergistic interpretability framework was implemented, integrating Pearson correlation, feature importance screening, SHapley Additive exPlanations (SHAP), and individual conditional expectation (ICE) plots. This multi-level approach quantitatively identified reinforcement characteristics, hot rolling reduction, and sintering pressure as the primary mechanical descriptors. Crucially, the model captured distinct, non-linear optimization regimes, identifying optimal property windows at lower pressures (40–50 MPa) for spark plasma sintering (SPS) and higher pressures (300–500 MPa) for hot-pressing, both in synergy with 70–75% rolling reductions. Experimental validation on a 2.0 wt% MXene/Ti composite confirmed the model's high-fidelity strength prediction with errors <2%. This ML-driven strategy provides a computationally efficient paradigm for defining tailored processing envelopes, offering new insights into the inverse design of high-performance metal matrix composites.

AB - The design of high-performance titanium matrix composites (TMCs) often faces the challenge of complex, nonlinear interactions between multi-scale processing parameters and mechanical properties. To overcome the limitations of traditional empirical methods, this study establishes a comprehensive data-driven computational framework to predict and optimize the performance of TMCs. To address the challenge of small-sample generalization, a diversity-oriented independent validation strategy was employed, strictly separating model training from a heterogeneous hold-out test set. Among various evaluated algorithms, the gradient boosting (GB) model demonstrated superior fidelity, achieving high predictive accuracy on the unseen independent datasets (R2 up to 0.88 for strength). To ensure model transparency, a synergistic interpretability framework was implemented, integrating Pearson correlation, feature importance screening, SHapley Additive exPlanations (SHAP), and individual conditional expectation (ICE) plots. This multi-level approach quantitatively identified reinforcement characteristics, hot rolling reduction, and sintering pressure as the primary mechanical descriptors. Crucially, the model captured distinct, non-linear optimization regimes, identifying optimal property windows at lower pressures (40–50 MPa) for spark plasma sintering (SPS) and higher pressures (300–500 MPa) for hot-pressing, both in synergy with 70–75% rolling reductions. Experimental validation on a 2.0 wt% MXene/Ti composite confirmed the model's high-fidelity strength prediction with errors <2%. This ML-driven strategy provides a computationally efficient paradigm for defining tailored processing envelopes, offering new insights into the inverse design of high-performance metal matrix composites.

KW - Machine learning

KW - Processing-property relationship

KW - SHAP analysis

KW - Small-data learning

KW - Titanium matrix composites

UR - https://www.scopus.com/pages/publications/105033238934

U2 - 10.1016/j.commatsci.2026.114624

DO - 10.1016/j.commatsci.2026.114624

M3 - Article

AN - SCOPUS:105033238934

SN - 0927-0256

VL - 268

JO - Computational Materials Science

JF - Computational Materials Science

M1 - 114624

ER -