Scholarlink

<p>In this study, experimental measurements of surface roughness were conducted on samples produced using Laser Powder Bed Fusion, with particular attention to the influence of scanning speed and laser power in both the body and contour regions. The roughness was evaluated at different heights within the build volume and across multiple surface orientations relative to the recoater blade. Based on the collected data, a feedforward Artificial Neural Network was trained using process parameters and geometrical factors (height and angle) as input, and experimentally measured roughness as output. The trained model achieved high predictive accuracy, with a Pearson correlation coefficient of 0.92 between predicted and measured values. Using the predictive capabilities of the Artificial Neural Network, it was possible to identify optimized combinations of process parameters for each combination of height and angle within the build volume. The optimization led to a significant reduction in surface roughness, lowering the minimum measured <i>Ra</i> (at Height = 15mm and Angle = 180°) from 5.51µm (experimentally obtained) to a predicted 3.46µm under the optimal parameter configuration, and achieving an average reduction of approximately 19.4% across the other positions.</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

Predictive modeling and optimization of surface roughness in LPBF through artificial neural networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月5日

查看全文
  • Delio Lusicini,
  • Matteo Crachi,
  • Raffaella Sesana,
  • Cristiana Delprete,
  • Marco Pizzarelli,
  • Nicola Sicignano,
  • Domenico Borrelli,
  • Antonio Caraviello

摘要

In this study, experimental measurements of surface roughness were conducted on samples produced using Laser Powder Bed Fusion, with particular attention to the influence of scanning speed and laser power in both the body and contour regions. The roughness was evaluated at different heights within the build volume and across multiple surface orientations relative to the recoater blade. Based on the collected data, a feedforward Artificial Neural Network was trained using process parameters and geometrical factors (height and angle) as input, and experimentally measured roughness as output. The trained model achieved high predictive accuracy, with a Pearson correlation coefficient of 0.92 between predicted and measured values. Using the predictive capabilities of the Artificial Neural Network, it was possible to identify optimized combinations of process parameters for each combination of height and angle within the build volume. The optimization led to a significant reduction in surface roughness, lowering the minimum measured Ra (at Height = 15mm and Angle = 180°) from 5.51µm (experimentally obtained) to a predicted 3.46µm under the optimal parameter configuration, and achieving an average reduction of approximately 19.4% across the other positions.