Scholarlink

<p>Al-Ti dissimilar welding enables lightweight, high-performance multi-material structures but is fundamentally constrained by the rapid formation of brittle intermetallic compounds (IMCs) at the interface, which severely compromise joint integrity. Conventional numerical approaches, including FEM/CALPHAD coupling and phase-field modeling, effectively capture diffusion-driven IMC evolution but are computationally intensive and unsuitable for real-time process optimization. Conversely, purely data-driven machine learning methods often lack physical consistency and generalization capability. In this study, a multi-physics physics-informed neural network (PINN) framework is developed to predict interfacial microstructure evolution in Al-Ti joints by embedding governing physics of multicomponent diffusion with Kirkendall effect, phase-field kinetics, solute trapping under rapid solidification, and anisotropic interfacial energy. The framework is validated using experimental microstructure datasets from laser beam welding, friction stir welding, and ultrasonic additive manufacturing of Al6061-Ti6Al4V joints. The proposed model achieves phase fraction prediction accuracy below 10<sup>−4</sup> mean squared error and an intermetallic layer thickness root-mean-square error of approximately 0.32&#xa0;μm, while delivering up to a 50 × reduction in computational time compared with conventional phase-field simulations. These results establish a data-efficient, physically consistent surrogate modeling approach for real-time microstructure prediction and adaptive control in Al-Ti dissimilar joining.</p> Graphical Abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AI-Driven Predictive Modeling of Interfacial Microstructure Evolution in Aluminum-Titanium Dissimilar Welds Using Physics-Informed Neural Networks

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Ankit Tyagi,
  • Abhishek Dadhich,
  • Ankita Dadhich,
  • Freedon Daniel

摘要

Al-Ti dissimilar welding enables lightweight, high-performance multi-material structures but is fundamentally constrained by the rapid formation of brittle intermetallic compounds (IMCs) at the interface, which severely compromise joint integrity. Conventional numerical approaches, including FEM/CALPHAD coupling and phase-field modeling, effectively capture diffusion-driven IMC evolution but are computationally intensive and unsuitable for real-time process optimization. Conversely, purely data-driven machine learning methods often lack physical consistency and generalization capability. In this study, a multi-physics physics-informed neural network (PINN) framework is developed to predict interfacial microstructure evolution in Al-Ti joints by embedding governing physics of multicomponent diffusion with Kirkendall effect, phase-field kinetics, solute trapping under rapid solidification, and anisotropic interfacial energy. The framework is validated using experimental microstructure datasets from laser beam welding, friction stir welding, and ultrasonic additive manufacturing of Al6061-Ti6Al4V joints. The proposed model achieves phase fraction prediction accuracy below 10−4 mean squared error and an intermetallic layer thickness root-mean-square error of approximately 0.32 μm, while delivering up to a 50 × reduction in computational time compared with conventional phase-field simulations. These results establish a data-efficient, physically consistent surrogate modeling approach for real-time microstructure prediction and adaptive control in Al-Ti dissimilar joining.

Graphical Abstract