Continuous high-throughput characterization of mechanical properties via deep learning

Abstract

High-throughput experiments (HTE) aim to acquire extensive chemical or physical properties in a single experiment, thereby enhancing testing efficiency. To simplify the extraction of diverse properties from one specimen, samples have moved from “discrete” arrays to “continuous” gradient ones. Despite this, complex responses of “continuous” gradient samples have impeded the development of continuous HTE. Full-field data, which can be obtained with Digital Image Correlation (DIC), is necessary for mechanical property characterizations. Traditional inversion methods for calculating property distributions from this data are slow and error-prone. Deep learning (DL) offers a faster and more accurate alternative for characterizing properties. Therefore, based on convolutional neural networks (CNNs), this article establishes a mapping model to obtain the modulus distribution directly from the full-field displacement. In view of the cost of time, simulation data are used to replace DIC data. However, fine mesh must be used to obtain the precise responses of gradient samples which unfortunately making the DL model face the challenge of time-consuming dataset generation and high-dimensional data mapping. To alleviate the difficulties, the isoparametric graded finite element (IGFE) formulation is introduced in this article, which offers an efficient way to generate datasets with low-dimension but high-fidelity. Results show that our framework not only has high prediction accuracy (with the L₁-error of 1.38%) but also enables fast characterization (within 12 ms), providing methodological support for high-throughput characterization based on gradient samples.