Counterexample-trained neural network model of rate and temperature dependent hardening with dynamic strain aging. (April 2022)

• Record Type: Journal Article
• Authors: Li, Xueyang; Roth, Christian C.; Bonatti, Colin; Mohr, Dirk
• Is Part Of: International journal of plasticity. Volume 151(2022)
• Publication Date: 2022-04
• Subjects: Machine learning -- Rate- and temperature effect -- Finite element analysis -- Dynamic strain aging -- Partial monotonicity in neural network -- Fracture initiation; Plasticity -- Periodicals; Plasticité -- Périodiques; Plasticity; Periodicals; 620.11233
• DOI: 10.1016/j.ijplas.2022.103218
• Languages: English
• ISSNs: 0749-6419

Abstract

Highlights: Performed more than 50 plasticity and fracture experiments (uniaxial, shear, notched tension, strain rate jump tests). Experiments cover strain rates of up to 150/s (before necking) and temperatures of up to 500 °C. Neural network function predicts hardening based on equivalent plastic strain, strain rate, temperature and a McCormick type of dynamic strain aging variable. Introduced counterexample-guided training to control strain rate effect. Extracted loading paths to fracture reveal a non-monotonic effect of the strain rate and the temperature. Identified a stress-state, strain rate and temperature dependent fracture initiation model. Abstract: Constitutive models dealing with the thermal and visco-plasticity of metals have seen wide applications in the automotive industry. A basic plasticity and fracture characterization of a 1.5 mm thick DP780 dual phase steel sheet based on uniaxial tensile (UT) experiments with seven distinct material orientations is complemented by low (∼0.001/s), intermediate (∼1/s) and high (∼150/s) strain rate experiments on notched tensile (NT) and shear (SH) specimens at temperatures ranging from 20 °C to 500 °C. At low strain rates, we observe a non-monotonic effect of the temperature on the force-displacement curves, with the highest curve obtained at 300 °C. Contrasting low speed tests, a monotonic effect of temperature is observed for intermediate and high strain rate experiments, with the highest curves obtained at 20 °C for both cases. Strain rate jump tests are performed proving the positive strain rate sensitivity of the steel. A machine-learning based plasticity model is developed to capture the observed complex strain rate- and temperature effect. The material is modeled as elasto-plastic, with a Hill'48 yield surface and a non-associated flow rule. The flow resistance is decoupled into a reference strain hardening term and a neural network term, which is a function of the plastic strain, strain rate, temperature, and an additional dynamic strain aging variable. The plasticity model is implemented into a material user subroutine and identified using a counterexample-guided hybrid experimental-numerical approach. The extracted loading paths reveal a complex rate and temperature effect on the ductility of DP780. A neural network based fracture initiation model is therefore adopted to describe the fracture onset across various stress states, strain rates and temperatures considered.