Physics-informed graph neural networks accelerating microneedle simulations towards novelty of micro-nano scale materials discovery

1 School of Integrated Science and Innovation (ISI), Sirindhorn International Institute of Technology (SIIT), Thammasat University, Thailand
2 Taiwan Semiconductor Research Institute National Applied Research Laboratories, Hsinchu 30091, Taiwan
3 National Electronics and Computer Technology Center (NECTEC), Pathum Thani 12120, Thailand

*Coresponding authors
The link for our paper will be freely open until Oct 27, 2023. Thanks to Elsevier's Share Link .
overview-of-the-work

This figure is the overview of our work. We propose a novel method combining between materials simulation and machine learning to accelerate the process of microneedle materials discovery.

Abstract

To commercialize the micro-nano needle as a drug delivery or biosensing platform, it is essential to validate the effectiveness and reliability of the piercing process for each needle material composition. However, proper microneedle material selection and fabrication requires sophisticated and costly semiconductor technology. This work aims to accelerate the material selection process of microneedles, focusing on polymeric materials due to their biocompatibility and flexibility. In this study, simulations of microneedles with various materials are used to generate training and testing data for a physics-informed machine learning model to predict von Mises stress distribution on microneedles of new materials. The training dataset is comprised of results from fifteen different materials. Different machine learning models are used, such as traditional tree-based, neural network, point cloud network, and graph-based models. Random-index selection is utilized to reduce the required number of data points by an order of magnitude. The graph attention network model is the best-performing model for predicting the von Mises stress of microneedles, with a mean square error of 8.3 × 10^(-5) MPa. The resulting model only requires 7 milliseconds to evaluate a new microneedle material, significantly faster than practical fabrication in a laboratory. The models also successfully handle data with a decimal scale obtained from microstructure simulations and predict physical behavior such as stress curves.

BibTeX


@article{chumpu2023mnmatdis,
  title = {Physics-informed graph neural networks accelerating microneedle simulations towards novelty of micro-nano scale materials discovery},
  journal = {Engineering Applications of Artificial Intelligence},
  volume = {126},
  pages = {106894},
  year = {2023},
  issn = {0952-1976},
  doi = {https://doi.org/10.1016/j.engappai.2023.106894},
  url = {https://www.sciencedirect.com/science/article/pii/S0952197623010783},
  author = {Romrawin Chumpu and Chun-Lin Chu and Tanyakarn Treeratanaphitak and Sanparith Marukatat and Shu-Han Hsu},
  keywords = {Machine learning, Graph neural networks, Microneedle, Numerical simulation},
}