Flexible and interpretable generalization of self-evolving computational materials framework

Thumbnail Image
Journal Title
Journal ISSN
Volume Title
Saudi Digital Library
The recent innovations of computational material models by machine learning (ML) methods face formidable challenges. Incorporating internal heterogeneity and diverse boundary conditions (BC’s) into existing ML methods remains difficult, and the weak interpretability of ML remains unresolved. To tackle these challenges, this dissertation generalizes a recently developed self-evolving computational material models framework built upon Bayesian update and evolutionary algorithm. This dissertation proposes a new material-specific information index (II), which is capable of autonomously quantifying the internal heterogeneity and diverse BC’s. Also, this dissertation introduces highly flexible cubic regression spline (CRS)-based link functions which can offer mathematical expressions of salient material coefficients of the existing computational material models in terms of convolved II. Thereby, this dissertation suggests a novel means by which ML can directly leverage internal heterogeneity and diverse BC’s to autonomously evolve computational material models while keeping interpretability. Validations using a wide spectrum of large-scale reinforced composite structures confirm the favorable performance of the generalization. Example expansions of nonlinear shear of quasi-brittle materials and progressive compressive buckling of reinforcing steel underpin efficiency and accuracy of the generalization. This dissertation adds a meaningful avenue for accelerating the fusion of computational material models and ML.