Nucleate boiling heat transfer enhancement using active controlled base plate trained through deep reinforcement learning and coarse grid simulations

We demonstrate deep reinforcement learning (DRL) for two problems. We first study drag reduction for flow past circular cylinder using transverse jets whose mass flow rate is controlled by artificial neural network. During training of the deep reinforcement learning agent, which is the jets in this case, we use coarser grid simulations for faster computations. We show that the model trained using coarser grid simulations works well with simulations at finer grid, thus saving significant computational time during the training phase. We next study the problem of nucleate boiling heat transfer enhancement at low superheats, with base plate actively controlled via DRL. Considering that Direct Numerical Simulations of nucleate boiling requires very fine grid size, during training of the DRL agent, we again perform coarser grid simulations, and then apply the learned model to finer grid independent simulations. The DRL agent shows good performance by trying to control the size and departure time of bubbes, leading to increase in convection, decrease in thermal boundary layer thickness and significant enhancement in Nusselt number.