Material development and optimization inherits the need to understand and control processing-structure-property relationships. Of particular interest is SiC, which is utilized in power electronic devices due to its exceptional properties such as a wide bandgap, high breakdown voltage, and high thermal conductivity [1]. The most common method to produce SiC single crystals is through physical vapour transport (PVT) where the SiC boules are grown within a closed reactor. Process temperatures around 2000 °C as well as pressure close to 5 mbar make it not only challenging to acquire experimental data, but also complicates the quantification of physical material properties of the system. In order to improve process knowledge and to reduce the need for expensive experimental trials, crystal growth simulations are frequently utilized. Reliable models not only facilitate the production of high-quality crystals as well as high output and resource-saving processes, but they can also aid in optimizing growth conditions or scaling the process appropriately. However, models must consider complex coupled physics phenomena, accurately representing the key aspects of the growth environment within reasonable computational efforts. A further challenge arises from uncertainty in the materials parameters of the furnace which need to be calibrated precisely to achieve reliable simulation results. Novel techniques from machine learning (ML) offer great opportunities in this respect. We present an approach providing a computational and time efficient active learning algorithm, that sequentially builds a ML model that can replace the COMSOL multiphysics finite element method (FEM) calculation to access relevant quantities in the reactor such as the temperature. Training data are generated by sequential Design of Experiment (DoE) leveraging Gaussian Process Regression (GPR). The COMSOL calculations are carried out sequentially and simultaneously the GPR model is utilized to determine the next set of parameters that would reduce the uncertainty of the ML model the most. The workflow is illustrated in Fig. 1. As a result of the absence of exact values for the physical material properties, input parameters for the ML model not only include the process parameters itself, but also material parameters and coefficients representing uncertainties, e.g. for radiation emissivities, thermal or electrical conductivities. These parameters are varied in a predefined range established by prior knowledge. The ML model was trained for one location close to the seed crystal, which can be measured experimentally. After gathering the training data through sequential DoE, ML models in combination with Bayesian inference, particularly with Markov Chain Monte Carlo (MCMC), are used to calibrate the unknown physical material parameters. Fig. 2. (a) shows the experimentally measured temperatures and the simulation results of the calibrated PVT simulation. The MCMC sample distribution for one parameter is shown in Fig. 2. (b). Furthermore, due to the high accuracy of this ML model, it can be used for different downstream tasks, like sensitivity and uncertainty analyses. Further potential applications and benefits of this approach in crystal growth facilities will be discussed.