In the industrial context, the fine modeling of radiative heat transfer is of paramount importance. For engine manufacturers, this transfer mode is predominant and determines the overall design of the product (wall thickness, pollutant production, etc.). This crucial information is also used to predict the life span of the system. We propose here a numerical methodology based on Monte Carlo method for solving the radiative transfer equation associated with Bayesian artificial neural networks to efficiently model the radiative flux divergence field in a highly heterogeneous axisymmetric domain. This work follows a first validation of the method on homogeneous cases [?]. A very intense and pitted hot spot was also added to the temperature and radiative species fields. This very punctual perturbation aims to demonstrate that the neural network is able to take into account all the heterogeneities of the study case and a strong irregularity often encountered in an industrial context (e.g. knocking phenomenon). Under the condition of a correctly constructed training database, it is possible to model such phenomena very accurately by a multivariate interpolation thanks to adapted neural networks. The use of these tools also allows to drastically reduce the computational time and the amount of resources mobilized for the resolution of the radiative heat transfer compared to a complete point by point calculation with a Monte Carlo method.