Scholarlink

<p>Geometric deviations in metal additive manufacturing (AM) pose a significant challenge to intelligent process planning and first-time-right fabrication, as their prediction and mitigation traditionally depend on computationally expensive finite element (FE) simulations or extensive experimental datasets. This study proposes a physics-informed intelligent manufacturing framework that integrates PointNet-based point cloud learning with gradient-based, decision-driven optimization to predict and compensate for such deviations. Unlike purely data-driven approaches, the framework is trained on validated thermo-mechanical FE simulations, embedding process physics into the learning process while requiring only minimal experimental data for calibration and validation. The PointNet architecture operates directly on raw 3D point cloud data, eliminating voxelization or meshing and preserving geometric fidelity through efficient, permutation-invariant learning. The proposed framework achieves high predictive accuracy (R<sup>2</sup>≥0.90) across representative cube and ring geometries and demonstrates robustness to in-plane rotational variations. Computational time is reduced from approximately 25&#xa0;min for FE simulation to under one minute for model inference, enabling near–real-time deployment. Furthermore, the iterative backpropagation-based compensation strategy functions as an autonomous geometric decision mechanism, reducing mean geometric error by an order of magnitude (from 0.1&#xa0;mm to 0.01&#xa0;mm), with over 99% of points lying within ± 5&#xa0;µm tolerance. By leveraging physics-based simulations instead of extensive experimental datasets, the proposed approach achieves good generalizability within topology class while minimizing experimental effort, establishing a scalable foundation for intelligent, decision-oriented additive manufacturing systems.</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

Physics-informed intelligent framework for geometric prediction and compensation in metal additive manufacturing

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • A. Nihil Mathew,
  • V. Pandurangan,
  • N. Venkaiah

摘要

Geometric deviations in metal additive manufacturing (AM) pose a significant challenge to intelligent process planning and first-time-right fabrication, as their prediction and mitigation traditionally depend on computationally expensive finite element (FE) simulations or extensive experimental datasets. This study proposes a physics-informed intelligent manufacturing framework that integrates PointNet-based point cloud learning with gradient-based, decision-driven optimization to predict and compensate for such deviations. Unlike purely data-driven approaches, the framework is trained on validated thermo-mechanical FE simulations, embedding process physics into the learning process while requiring only minimal experimental data for calibration and validation. The PointNet architecture operates directly on raw 3D point cloud data, eliminating voxelization or meshing and preserving geometric fidelity through efficient, permutation-invariant learning. The proposed framework achieves high predictive accuracy (R2≥0.90) across representative cube and ring geometries and demonstrates robustness to in-plane rotational variations. Computational time is reduced from approximately 25 min for FE simulation to under one minute for model inference, enabling near–real-time deployment. Furthermore, the iterative backpropagation-based compensation strategy functions as an autonomous geometric decision mechanism, reducing mean geometric error by an order of magnitude (from 0.1 mm to 0.01 mm), with over 99% of points lying within ± 5 µm tolerance. By leveraging physics-based simulations instead of extensive experimental datasets, the proposed approach achieves good generalizability within topology class while minimizing experimental effort, establishing a scalable foundation for intelligent, decision-oriented additive manufacturing systems.