<p>Corner cracks in continuously cast slabs are closely related to the high-temperature thermo-mechanical response of equiaxed austenite near slab corners, where intragranular slip interacts with grain-boundary embrittlement and damage. While coupled crystal plasticity finite element and cohesive zone (CPFEM–CZM) modeling can represent this competition, inverse calibration of its microscale parameters at casting-relevant temperatures is hindered by limited direct measurements, strong parameter coupling, and the high cost and convergence sensitivity of cohesive RVE simulations. This study develops a machine learning-assisted inverse calibration workflow at 940 °C to identify six coupled micro-parameters <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\left({\tau}_{0},{\tau}_{s},{h}_{0},{K}_{n},{\sigma}_{c},{\delta}_{c}\right)$</EquationSource> <EquationSource Format="MATHML"><math> <mfenced close=")" open="("> <msub> <mi>τ</mi> <mn>0</mn> </msub> <mo>,</mo> <msub> <mi>τ</mi> <mi>s</mi> </msub> <mo>,</mo> <msub> <mi>h</mi> <mn>0</mn> </msub> <mo>,</mo> <msub> <mi>K</mi> <mi>n</mi> </msub> <mo>,</mo> <msub> <mi>σ</mi> <mi>c</mi> </msub> <mo>,</mo> <msub> <mi>δ</mi> <mi>c</mi> </msub> </mfenced> </math></EquationSource> </InlineEquation>. A statistically representative 500-grain RVE is adopted based on isotropy convergence (CV = 1.42 pct). An initial database of 500 CPFEM–CZM simulations is generated within a physically admissible domain, and five stress–strain descriptors (yield/peak points and softening slope) are used as calibration targets. Two surrogates are trained (multi-output for pre-peak descriptors; single-output for the softening slope using ≈ 400 convergent curves), and Sobol-guided enrichment expands the database to 800 simulations. A hybrid DE-L-BFGS-B optimizer is applied and the identified parameters are validated by back-substitution CPFEM–CZM simulations, achieving a terminal full-curve MAPE of 0.98 pct and reducing the softening-regime error from 16.50 to 1.12 pct after enrichment.</p>

Embedded Machine Learning-Assisted Inverse Calibration of Crystal Plasticity and Cohesive Zone Constitutive Parameters for High-Temperature Equiaxed Austenite

期刊论文
发表日期：

2026年5月19日

查看全文

Junlong Ju,
Zhida Zhang,
Hai Chang,
Cheng Ji,
Miaoyong Zhu

摘要

Corner cracks in continuously cast slabs are closely related to the high-temperature thermo-mechanical response of equiaxed austenite near slab corners, where intragranular slip interacts with grain-boundary embrittlement and damage. While coupled crystal plasticity finite element and cohesive zone (CPFEM–CZM) modeling can represent this competition, inverse calibration of its microscale parameters at casting-relevant temperatures is hindered by limited direct measurements, strong parameter coupling, and the high cost and convergence sensitivity of cohesive RVE simulations. This study develops a machine learning-assisted inverse calibration workflow at 940 °C to identify six coupled micro-parameters $\left({\tau}_{0},{\tau}_{s},{h}_{0},{K}_{n},{\sigma}_{c},{\delta}_{c}\right)$ $(τ_{0}, τ_{s}, h_{0}, K_{n}, σ_{c}, δ_{c})$ . A statistically representative 500-grain RVE is adopted based on isotropy convergence (CV = 1.42 pct). An initial database of 500 CPFEM–CZM simulations is generated within a physically admissible domain, and five stress–strain descriptors (yield/peak points and softening slope) are used as calibration targets. Two surrogates are trained (multi-output for pre-peak descriptors; single-output for the softening slope using ≈ 400 convergent curves), and Sobol-guided enrichment expands the database to 800 simulations. A hybrid DE-L-BFGS-B optimizer is applied and the identified parameters are validated by back-substitution CPFEM–CZM simulations, achieving a terminal full-curve MAPE of 0.98 pct and reducing the softening-regime error from 16.50 to 1.12 pct after enrichment.