Scholarlink

<p>Additive manufacturing offers unparalleled freedom in designing complex structures but requires precise control over process parameters to ensure high-quality components. We have employed machine learning techniques to navigate between the parameters of laser-powder bed fusion (L-PBF) processes and predict wall thicknesses in the range of a few hundred microns (from 100 to 450 µm). Unsupervised learning has enabled the identification of the discrete process domains. Dimensional analysis and linear fits produced naive thickness predictions, albeit with limited success. Purely data-driven and physically informed Gaussian processes have achieved accurate maps of the process window. The informed models make use of physical relationships between dimensionless numbers, leading to machine- and material-independent predictions. Our results show that the physically informed Gaussian processes outperform the other methods, attaining a thickness prediction error 20% lower than pure data-driven approaches, and requiring fewer training points to reach similar performance levels. This study highlights the significance of combining the physical information of the system with machine learning approaches in order to master additive manufacturing processes.</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

Physically informed machine learning for the control of L-PBF processes

  • 期刊论文

  • 发表日期：

    2025年4月15日

查看全文
  • Antonio Peña Corredor,
  • Clément Ernould,
  • Jean-Marc Auger,
  • Ludovic Barrière

摘要

Additive manufacturing offers unparalleled freedom in designing complex structures but requires precise control over process parameters to ensure high-quality components. We have employed machine learning techniques to navigate between the parameters of laser-powder bed fusion (L-PBF) processes and predict wall thicknesses in the range of a few hundred microns (from 100 to 450 µm). Unsupervised learning has enabled the identification of the discrete process domains. Dimensional analysis and linear fits produced naive thickness predictions, albeit with limited success. Purely data-driven and physically informed Gaussian processes have achieved accurate maps of the process window. The informed models make use of physical relationships between dimensionless numbers, leading to machine- and material-independent predictions. Our results show that the physically informed Gaussian processes outperform the other methods, attaining a thickness prediction error 20% lower than pure data-driven approaches, and requiring fewer training points to reach similar performance levels. This study highlights the significance of combining the physical information of the system with machine learning approaches in order to master additive manufacturing processes.