Scholarlink

We address the problem of insufficient data for neural network regression modelling in the context of mechanical properties of 3D printed scaffolds. Experimental data corresponds to compressive strength of PLA scaffolds dependent of layer thickness, extrusion temperature and porosity. This paper aims to evaluate two data augmentation strategies based on a Gaussian Copula (GC) method and on a conditional tabular generative adversarial network (CTGAN) method. The results indicate that the GC method generates a more statistically similar data than the CTGAN method. A feed-forward neural network with fixed 3-8-1 architecture and Levenberg-Marquardt learning algorithm is used for regression using synthetic and real data from both methods and evaluated through the coefficient of determination (R) and mean square error (MSE). Both metrics point to a better regression performance of the GC method, with R-values above 80% and lower MSE values compared to lower R-values (less than 40%) and larger MSE values of the CTGAN method. The performance plots show a better stability of the network’s training-validation processes for data generated with the GC method.

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

Data Augmentation Strategies for Machine Learning Modelling of Compressive Strength of Biomedical Scaffolds

  • 章节

  • 发表日期：

    2026年

查看全文
  • Alejandro González González,
  • Patricia Zambrano-Robledo,
  • Rolando Javier Lima de la Torre,
  • Sandra García Ortega

摘要

We address the problem of insufficient data for neural network regression modelling in the context of mechanical properties of 3D printed scaffolds. Experimental data corresponds to compressive strength of PLA scaffolds dependent of layer thickness, extrusion temperature and porosity. This paper aims to evaluate two data augmentation strategies based on a Gaussian Copula (GC) method and on a conditional tabular generative adversarial network (CTGAN) method. The results indicate that the GC method generates a more statistically similar data than the CTGAN method. A feed-forward neural network with fixed 3-8-1 architecture and Levenberg-Marquardt learning algorithm is used for regression using synthetic and real data from both methods and evaluated through the coefficient of determination (R) and mean square error (MSE). Both metrics point to a better regression performance of the GC method, with R-values above 80% and lower MSE values compared to lower R-values (less than 40%) and larger MSE values of the CTGAN method. The performance plots show a better stability of the network’s training-validation processes for data generated with the GC method.