Scholarlink

<p>Defects in metal additive manufacturing (AM) are closely related to multiphysics fields. However, large-scale thermo-mechanical simulations used to capture this multiphysics behavior are computationally expensive and time-consuming. This paper proposes AMBayes, a Bayesian data-driven framework that integrates physical modeling with efficient full-field prediction and process optimization. AMBayes learns a feature-based Gaussian Process (GP) surrogate model of large-scale spatiotemporal multiphysics responses from limited samples. A real-time communication pipeline between the surrogate model and Bayesian Optimization (BO) allows accurate objective evaluations in seconds, enabling rapid posterior updates and advancing the optimization process. On a challenging laser-directed energy deposition (L-DED) optimization problem, AMBayes mitigates harmful residual stress and improves temperature uniformity in 15.97&#xa0;s, while providing full-field predictions. Furthermore, the data-driven approach requires only a very small dataset (e.g., 32 samples), considerably lowering the effort required to collect expensive large-scale simulation data.</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

AMBayes: a Bayesian data-driven framework for large-scale full-field prediction and process optimization with limited samples in additive manufacturing

  • 期刊论文

  • 发表日期：

    2025年12月8日

查看全文
  • Haoyang Luo,
  • Chufan He,
  • Huimin Xie,
  • Wei He,
  • Chensen Ding

摘要

Defects in metal additive manufacturing (AM) are closely related to multiphysics fields. However, large-scale thermo-mechanical simulations used to capture this multiphysics behavior are computationally expensive and time-consuming. This paper proposes AMBayes, a Bayesian data-driven framework that integrates physical modeling with efficient full-field prediction and process optimization. AMBayes learns a feature-based Gaussian Process (GP) surrogate model of large-scale spatiotemporal multiphysics responses from limited samples. A real-time communication pipeline between the surrogate model and Bayesian Optimization (BO) allows accurate objective evaluations in seconds, enabling rapid posterior updates and advancing the optimization process. On a challenging laser-directed energy deposition (L-DED) optimization problem, AMBayes mitigates harmful residual stress and improves temperature uniformity in 15.97 s, while providing full-field predictions. Furthermore, the data-driven approach requires only a very small dataset (e.g., 32 samples), considerably lowering the effort required to collect expensive large-scale simulation data.