In combustion simulation, a computation of detailed chemistry can cost over 90 % of simulation time. Tabulated chemistry offers an established technique by trading off reduced runtime for  increasing memory requirements. Depending on the complexity of the chemistry model, the resulting database can exceed  manageable volume. A smart replacement of the table can circumvent this problem. This thesis investigates different machine learning methods for modeling thermochemical manifolds using a simple 2D Flamelet Generated Manifold (FGM) as an example case.  It is examined how Principle Component Analysis (PCA) can parameterize thermochemical manifolds. Additionally, regression algorithms are analyzed and an Artificial Neural Network (ANN) is developed capturing a 2D FGM excellently. For that, different modeling aspects are studied which give a guideline for the design of ANNs in the context of FGM representation. Furthermore, a clustering algorithm grouping thermochemical scalars is developed which can be applied to reduce the complexity of ANNs while  improving prediction quality. Finally, it is shown that transfer  learning reduces training time about an order of magnitude.