How to overcome limitations of the Abaqus superelasticity material model with reduced order modeling and machine learning
How to overcome limitations of the Abaqus superelasticity material model with reduced order modeling and machine learning
Tuesday, May 5, 2026
The accurate numerical representation of Nitinol’s superelastic behavior remains a key challenge for designing reliable medical devices. Existing constitutive models often excel in academic studies but fall short in practical engineering workflows, where computational efficiency, robustness, and parameter accessibility are essential.
This contribution discusses the steps towards an industry-affine material model for Nitinol and evaluates which physical effects must be captured to ensure predictive capability, and which complexities may be omitted without significant loss of accuracy. The commonly used superelastic material formulation implemented in Abaqus for structural FEM computations is based on the models of Auricchio and Taylor (1996, 1997) and serves as a reference. Despite its widespread use, this model exhibits several limitations: (1) the compression regime is modeled in a simplified way (2) residual elongation after unloading due to remaining martensite is not captured (3) unloading after plastic deformation does not match with experimental data (4) initial superelastic anisotropies due to manufacturing processes are neglected, to name some of the most industry-relevant ones.
Then, two solution strategies are presented: (i) model order reduction approach for complex material models and (ii) solving superelasticity by Constitutive Artificial Neural Networks (CANNs). Finally, the long-term perspective of automated model discovery for Nitinol is addressed.
This contribution discusses the steps towards an industry-affine material model for Nitinol and evaluates which physical effects must be captured to ensure predictive capability, and which complexities may be omitted without significant loss of accuracy. The commonly used superelastic material formulation implemented in Abaqus for structural FEM computations is based on the models of Auricchio and Taylor (1996, 1997) and serves as a reference. Despite its widespread use, this model exhibits several limitations: (1) the compression regime is modeled in a simplified way (2) residual elongation after unloading due to remaining martensite is not captured (3) unloading after plastic deformation does not match with experimental data (4) initial superelastic anisotropies due to manufacturing processes are neglected, to name some of the most industry-relevant ones.
Then, two solution strategies are presented: (i) model order reduction approach for complex material models and (ii) solving superelasticity by Constitutive Artificial Neural Networks (CANNs). Finally, the long-term perspective of automated model discovery for Nitinol is addressed.