An artificial intelligence (AI) enhanced optimization framework is developed to reduce computational costs for evaluating transport performance of buoyancy driven heat and mass transfer in porous structures. The present optimization framework integrates prediction with artificial neural networks (ANNs), optimization with the weighted objective function, and physics-based simulations with high performance computing (HPC). Multi-dimensional governing parameters and objectives are investigated by ANNs with sparse scattered training data obtained from HPC with controllable structure generation scheme (CSGS) and parallel non-dimensional lattice Boltzmann method (P-NDLBM). The macroscopic prediction results based on ANNs are validated by comparison with HPC results. Full maps of the objective function values versus structure and physical parameters are illustrated. The maximum objective function value subjected to constraints is obtained together with the corresponding optimal structure and physical parameters. The optimal parameters are further applied in HPC to obtain mesoscopic physical fields. The underlying mechanism is also revealed by comparing the physical fields with optimal and off-optimal parameters.