Integrating ML-Based Models into the INSANE Finite Element System

Resumo

Machine learning (ML) models have emerged as promising alternatives to classical constitutive formulations for representing quasi brittle materials, such as concrete, rock, and composite media. Unlike traditional analytical approaches, ML models are trained directly on experimental or simulated data, enabling them to capture complex nonlinear behaviors—such as stiffness degradation, progressive cracking, and brittle-to-ductile transitions—in a more adaptive and generalizable manner. Their inherent flexibility allows for the incorporation of material variability and the customization of constitutive responses according to boundary conditions and loading history.However, the integration of ML-based constitutive models into finite element simulation environments demands substantial modifications to the underlying software architecture. Since such models operate through predictions made by neural networks or similar algorithms, they require dedicated structures for storing and updating internal variables, as well as efficient communication mechanisms with the numerical core to ensure the stability and performance of the analysis.This work presents the strategies adopted to adapt the INSANE system—originally developed in Java—to support ML-based constitutive modeling. Interoperability mechanisms with Python environments were implemented, alongside robust runtime data exchange interfaces and modifications to the constitutive integration logic, while preserving the modularity of the system. The adopted approach enables future extensions to various ML architectures and enhances INSANE’s predictive capabilities in simulations involving brittle and heterogeneous materials.