Machine Learning-Assisted Inverse Design and Finite Element Validation of Auxetic Structures for Biomedical Applications


Balcı H. Ş., Balcı F.

ULUSLARARASI YÜKSEK İHTİSAS ÜNİVERSİTESİ SAĞLIK BİLİMLERİ KONGRESİ, Ankara, Türkiye, 16 - 18 Ekim 2026, ss.1, (Özet Bildiri)

  • Yayın Türü: Bildiri / Özet Bildiri
  • Basıldığı Şehir: Ankara
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.1
  • Gazi Üniversitesi Adresli: Evet

Özet

Auxetic structures exhibit mechanical behaviors that differ from those of conventional materials, such as transverse expansion under tensile loading and transverse contraction under compressive loading, owing to their negative Poisson’s ratios. These characteristics offer significant design potential for customizing biomedical implants, tissue engineering structures, and mechanical meta-structures intended for long-term use according to patient- and application-specific requirements. In this study, a machine learning-based inverse design approach was developed to enable the automatic dimensioning of auxetic unit cells with a star geometry according to a target Poisson’s ratio. While the outer frame of the designed unit cell varied within the range of 2000–3000 µm, the wall thickness was kept constant at 100 µm. Different regression models were compared using the dataset obtained from finite element analyses, and radial basis function kernel support vector regression (RBF-SVR) was identified as the most suitable model. The model hyperparameters were determined using Tree-structured Parzen Estimator-based Bayesian optimization, and the trained forward prediction model was used for inverse design within the continuous geometric design space.

The method was applied to six target Poisson’s ratios ranging from -0.20 to -0.70, and a total of 30 new designs were generated, with five different geometries for each target value. In the validation performed using finite element analyses, R² = 0.9830, MAE = 0.01584, RMSE = 0.02226, and MAPE = 3.39% were obtained, while the mean deviation was calculated as -0.00246. In 93.3% of the validation samples, the absolute error remained below 0.05.

The results demonstrate that micrometer-scale auxetic cells provide an investigable design framework for stent geometries, bone tissue scaffolds, and other implantable meta-structures. Directly translating the targeted mechanical property into geometry offers a sustainable approach by reducing repetitive finite element analyses and prototyping requirements, thereby enabling more efficient use of computational, material, and time resources.