Application of machine learning models for estimation of material properties

In the process of designing a part or a component, computer simulations are extensively used as they test the intended function of a product. Complex load conditions need to be adequately represented in simulations to make reliable estimations of the behavior of a structure. The material of a structure influences its mechanical behavior and is defined through various material models. The choice of the material model will depend on the conditions in which the structure will be used. After choosing a material model, user must define model parameters or use libraries with available material data of common materials. As not all material parameters can easily be acquired from experiments, data-driven models are used to find relationships between desired material parameters and the available ones. Data-driven models or, more specifically, machine learning (ML) models enable detecting patterns and extracting new relationships between material parameters which could not be discovered using the classical empirical models [1]. ML models need to be built on a sufficient amount of data, yet a common problem concerning material characterization is small low-dimensional data sets [2]. To overcome this problem, important input variables need to be detected by using feature selection techniques [3]. Furthermore, material properties from existing material databases need to be collected and used in building ML models to increase the number of learning examples [2]. These models should enable the end-user to efficiently design a product without the need to carry out time-consuming experimental material testing.