Simulation in the industrial design process continues to grow in importance. Underlying models become more and more complex because of different physical phenomena and additional models such as cost models. These comprehensive models are multi-scale models that require high numerical effort to solve. When solving optimization tasks, these complex models have to be solved several times, which causes additional computational costs. For this reason, surrogate models are derived that require significantly lower computing costs. Surrogate models can be black box models or gray box models. Grey-box models combine data-driven with physics-based models to improve the reliability of the model. Physics Informed Neural Networks (PINN) are used as grey-box models to approximate the behavior of dynamic systems. We consider stiff ordinary differential equations (ODEs) as examples for chemical reaction models. ODEs are stiff if the solution being sought is varying slowly, but there are nearby solutions that vary rapidly, so the numerical integration method must use small time steps to obtain a satisfactory accuracy. Numerical integration methods solve the ODEs step-wise. The integration error depends on the order of the scheme and on the type of the scheme (explicit and implicit). Implicit schemes can approximate larger time steps than explicit schemes, but often require high computational effort. If a time step is learned for arbitrary initial values, computational effort be can reduced under certain circumstances. We present a deep learning algorithm for stiff ODEs that obtains additional properties of the ODE, such as positivity, conservation and additional properties of auto-catalytic reactions. The algorithm is discussed on two examples. The implementation is done in \href{https://github.com/boschresearch/torchphysics}{TorchPhysics}, an open source code for the PINN method based on PyTorch and Python.