Machine learning meta-models for fast parameter identification of the lattice discrete particle model

When simulating the mechanical behavior of complex materials, the failure behavior is strongly influenced by the internal structure. To account for such dependence, models at the length scale of material heterogeneity are required. These models involve multiple material parameters and are computationally intensive. Experimental data are needed to identify model parameters, and the highly nonlinear nature of the constitutive equations results in a challenging inverse problem. Direct inverse analysis (DIA) seeks the best parameter estimates by minimizing a well-defined objective function through an iterative optimization scheme. However, it is time-consuming, as just a single simulation is computationally costly. Another approach uses a machine learning (ML) model built from the complete mechanistic model, combined with an appropriate optimization algorithm. ML reduces the computational cost and enables parameter selection and feature importance as a by-product. This manuscript presents a comparative study between DIA and ML-based inverse analysis using the lattice discrete particle model, a state-of-the-art model simulating concrete at the coarse aggregate level. The study focuses on three mechanical tests: unconfined compression, hydrostatic, and tensile fracture. Experimental data was taken from the literature and augmented to form a consistent data set for a given mix design. Five different ML methods were explored, and results were compared with those from DIA. The two inverse analysis methods were compared in terms of goodness of fit and computational cost. Results confirm the validity of the identification procedure and show that inverse analysis based on ML reduces the computational cost by various orders of magnitude.