Modelling heterogeneous materials requires knowledge about their microstructure. A fundamental way to represent material geometry on a microscopic level in functional form is a three-point statistical descriptor. Since there is no closed form representation for n-point correlation functions with n>2, the usual approach is an approximation using a combination of two-point functions or a direct extraction from a microstructure image. In the first case, the approximation does not have high accuracy, the second approach can be computational costly. Here, we present an ANN surrogate model for three-point correlation functions to achieve both accuracy and efficiency. For the approach, we use a combination of CNN and FCNN networks, based on the structure of neural operators. Furthermore, combining a purely geometric surrogate model with material specific modelling allows for high adaptability. The method is applicable to Carbon Fiber Reinforced Composites (CFRPs).