Sloshing suppression with a controlled elastic baffle via deep reinforcement learning and SPH simulation

This study employed smoothed particle hydrodynamics (SPH) as the numerical environment, integrated with deep reinforcement learning (DRL) real-time control algorithms to optimize sloshing suppression in a tank with a centrally positioned vertical elastic baffle. Compared with moving rigid-baffle control and active-strain control, the DRL-controlled moving elastic baffle achieved the best performance, reducing the mean free-surface amplitude by 81.63% in the baseline case. A cosine-based expert policy derived from the learned velocity profile was further evaluated in a representative three-dimensional (3D) simulation and achieved a 76.86% reduction, providing a preliminary 2D-to-3D transfer assessment under this condition rather than evidence of broad cross-dimensional transferability. Energy analyses showed that elastic baffle motion effectively decreased the system energy by performing negative work on the fluid, reducing both kinetic and potential energy. The tested transfer cases over different excitation frequencies and water depths also showed substantial suppression, but with reductions outside a single 70%–80% interval; for example, displacement-based strategies ranged from 64.85% to 81.63%, while the active-strain baseline reached 41.5%. These results indicate that DRL-informed moving-baffle control is effective in the tested nonlinear FSI conditions, while broader parametric training and direct 3D fine-tuning remain necessary future work.