Scholarlink

<p>Mechanical testing is a critical but often protracted process for evaluating materials. This study demonstrates a novel high-throughput approach that integrates femtosecond laser machining for rapid sample preparation, automated tensile testing of specimen arrays, and machine learning techniques for efficient data analysis. 316L stainless steel and additively manufactured grade 91 steel were used to fashion miniature tensile specimens. The mechanical properties were automatically extracted from the ensuing stress-strain curves using both a supervised deep learning segmentation model (U-Net) and unsupervised clustering methods (k-means, DBSCAN). While all techniques performed acceptably on the more homogeneous 316L samples, the trained U-Net showed superior robustness and accuracy when analyzing the highly heterogeneous grade 91 specimens, with errors 2–3 times lower than the unsupervised approaches compared to manual analysis. The initial expense incurred generating training data for the U-Net was offset by significantly decreased analysis time and improved consistency. This unified methodology, combining machining, automated testing, and machine learning, provides an accelerated workflow for investigating mechanical properties of both additively manufactured and conventional alloys.</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 Novel Machine Learning-Driven Approach to High Throughput Mechanical Testing

  • 期刊论文

  • 发表日期：

    2025年1月16日

查看全文
  • Tongjun Niu,
  • Ross Lee,
  • Sebastian Lam,
  • Joseph Hafen,
  • Steven Lukow,
  • Nan Li,
  • Peter Hosemann,
  • Jonathan Gigax

摘要

Mechanical testing is a critical but often protracted process for evaluating materials. This study demonstrates a novel high-throughput approach that integrates femtosecond laser machining for rapid sample preparation, automated tensile testing of specimen arrays, and machine learning techniques for efficient data analysis. 316L stainless steel and additively manufactured grade 91 steel were used to fashion miniature tensile specimens. The mechanical properties were automatically extracted from the ensuing stress-strain curves using both a supervised deep learning segmentation model (U-Net) and unsupervised clustering methods (k-means, DBSCAN). While all techniques performed acceptably on the more homogeneous 316L samples, the trained U-Net showed superior robustness and accuracy when analyzing the highly heterogeneous grade 91 specimens, with errors 2–3 times lower than the unsupervised approaches compared to manual analysis. The initial expense incurred generating training data for the U-Net was offset by significantly decreased analysis time and improved consistency. This unified methodology, combining machining, automated testing, and machine learning, provides an accelerated workflow for investigating mechanical properties of both additively manufactured and conventional alloys.