Heat-Transfer is a central phenomena for many process industries. It is also a the core of the energy transition since today about 2/3rd of the world primary energy use is not providing any service, often ending up as low temperature waste-heat. The quest for novel technologies for harvesting unused heat has led to new research and uses advanced computational tools. However, engineers are still not able to use these advanced high-fidelity tools and the lead-time is prohibitive in the product design phase. Therefore, there is a need for rapid tools that can rapidly predict heat fluxes. Our group at KTH has been building an extensive database for reacting heat-transfer problems using high-fidelity simulations. These simulations have evidenced the mechanisms and revealed the physics driving the heat-transfer intensification. A natural next step is to use this database to create data-driven Reduced-Order-Models (ROM) that will be of use for the engineering community. This project aims at creating a model based on beta variational autoencoder (VAE). We combine statistical descriptions of the data with deep neural networks, and enable introducing stochasticity to the latent space, where the low-dimensional representation of the data is obtained. By regularizing the loss function of the VAE, we will encourage learning of statistically-independent variables while penalizing the size of the latent space. The VAE will be combined with a transformer to advance in time the system without solving CFD equations. We will consider both hot and cold plate cases. The predictions will be evaluated and compared to the original data and to the linear modal predictions. The code for this work will be based on the article by Vinuesa's group that is already running on Berzelius. Reference: https://www.nature.com/articles/s41467-024-45578-4