For polymers, inelastic deformation and failure under dynamic loading is largely dependent on temperature, strain rate and stress multiaxiality. Consequently, modelling their impact behavior requires several experiments and time-consuming determination of material parameters. In this work, a calibration strategy to reduce calibration times and experimental workloads based on neural networks and genetic algorithms is presented. The method is demonstrated on Charpy- and puncture experiments at 23 \( ^{\circ }\text {C}\) and 50 \( ^{\circ }\text {C}\) with two additional tensile tests at \(-\) 30 \( ^{\circ }\text {C}\) and 23 \( ^{\circ }\text {C}\) . In many cases, these experiments are conducted for datasheet specifications and are readily available. To achieve a balance between precision, ability to generalize and numerical simplicity, the strain-rate and temperature dependence was mapped via a Johnson–Cook model that creates tabular inputs for a general J2 plasticity model in Abaqus. Similarly, failure behavior was modelled using ductile-damage in Abaqus. The traditional fitting methodology based on inverse analysis using finite element simulations was replaced by the use of neural network surrogate models which allow for calibration in a matter of minutes. This approach was validated against four-point bending experiments with geometry, velocity and temperatures different to those used to calibrate the model on. Here, simulated toughness and limit loads on average differed by 19% and 5% from experimental results.
