A virtual sample generation algorithm supporting machine learning with a small-sample dataset: A case study for rubber materials

Machine learning (ML) is widely used in the field of material informatics. However, limitations on the size of available datasets are a key bottleneck in the use of machine learning methods to predict material properties or reverse-design high-performance materials. To solve this problem, we propose a virtual sample generation algorithm based on a Gaussian mixture model (GMM-VSG) to address the lack of training samples in machine learning. The core idea of the algorithm is to generate virtual samples by fitting the distribution of the original samples. We used an open rubber composite dataset (24 samples) to establish a machine learning model to predict the wear resistance of rubber materials through mechanical properties to verify the performance of the GMM-VSG algorithm. The results show that after using our algorithm, the R2 of the prediction model reached 0.95, and the prediction accuracy increased by 41%. This shows that the proposed algorithm can effectively promote the prediction accuracy of data model with small sample size.