A mechanical property prediction system for G-Lattices via machine learning


Armanfar A., TAŞMEKTEPLİGİL A. A., Üstündağ E., Lazoglu I., Günpınar E.

Engineering Optimization, 2024 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Publication Date: 2024
  • Doi Number: 10.1080/0305215x.2023.2295353
  • Journal Name: Engineering Optimization
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Aerospace Database, Communication Abstracts, Compendex, Metadex, zbMATH, Civil Engineering Abstracts
  • Keywords: computer-aided design, G-Lattices, Lattice structure, machine learning, mechanical property
  • Istanbul Technical University Affiliated: Yes

Abstract

G-Lattices—a novel family of periodic lattice structures introduced by Arash Armanfar and Erkan Gunpinar—demonstrate diverse mechanical properties owing to their generatively designed shapes. To assess the properties of lattice structures effectively, experimental tests and finite element analysis (FEA) are commonly used. However, the complex nature of these structures poses challenges, leading to high computation time and costs. This study proposes a machine learning (ML) approach to predict the mechanical properties of G-Lattices quickly under defined loading conditions. G-Lattice training data is generated through a sampling technique, and voxelized data is employed as ML feature vectors for predicting properties determined by FEA. To address the uneven distribution of target values, samples are clustered and utilized to train a classification model. This two-step process involves the classification of G-Lattices, followed by the application of specific regression models trained for each cluster for precise predictions. According to experiments, the ML model obtained, which predicts stiffness-over-volume ratios for G-Lattices, achieved a mean absolute percentage error of 6.5% for 1600 G-Lattices in a few seconds. Furthermore, approximately 70% of the 40,000 G-Lattices exhibited errors within 5%. The ML model's rapid predictions and acceptable accuracy make it useful for quick decision-making and seamless integration into optimization processes.