Scholarlink

<p>The uncertainty in critical simulation parameters significantly limits the predictive accuracy of thermal-structural coupled models in laser-directed energy deposition (DED). This study develops an efficient calibration framework to determine unmeasurable parameters and improve finite element simulation fidelity. Traditional trial-and-error approaches are computationally prohibitive due to high simulation costs. To address this, we propose a machine learning-assisted multi-parameter calibration framework integrating physics-informed neural networks with energy conservation principles to construct a computationally efficient surrogate model. SHAP (SHapley Additive exPlanations) analysis identifies the most sensitive parameters requiring calibration, and the sparrow search algorithm optimizes these parameters by minimizing discrepancy between simulated and experimentally measured substrate temperatures. Experimental validation demonstrates that the calibrated model reduces temperature prediction error from 27.9 to 9.7%, achieving approximately 65% improvement. This study provides an effective and computationally efficient approach for simulation parameter calibration in additive manufacturing, enhancing numerical model reliability for engineering applications.</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

A physics-informed neural network framework for multi-parameter calibration of DED thermal models

  • 期刊论文

  • 发表日期：

    2026年3月5日

查看全文
  • Huayang Xiang,
  • Yuxiang Ji,
  • Zhenfei Zhan,
  • Zhao Liu,
  • Zexi Li,
  • Zhongjie Yue

摘要

The uncertainty in critical simulation parameters significantly limits the predictive accuracy of thermal-structural coupled models in laser-directed energy deposition (DED). This study develops an efficient calibration framework to determine unmeasurable parameters and improve finite element simulation fidelity. Traditional trial-and-error approaches are computationally prohibitive due to high simulation costs. To address this, we propose a machine learning-assisted multi-parameter calibration framework integrating physics-informed neural networks with energy conservation principles to construct a computationally efficient surrogate model. SHAP (SHapley Additive exPlanations) analysis identifies the most sensitive parameters requiring calibration, and the sparrow search algorithm optimizes these parameters by minimizing discrepancy between simulated and experimentally measured substrate temperatures. Experimental validation demonstrates that the calibrated model reduces temperature prediction error from 27.9 to 9.7%, achieving approximately 65% improvement. This study provides an effective and computationally efficient approach for simulation parameter calibration in additive manufacturing, enhancing numerical model reliability for engineering applications.

Graphical abstract