Nonlocal density descriptor for machine-learning-corrected density functional

The integration of artificial intelligence with density functional theory offers novel opportunities to advance the accuracy and efficiency for practical applications. However, current machine learning models based on semilocal density descriptors still face limitations in their ability to represent the exchange-correlation functional. In this work, we propose a novel approach for evaluating a nonlocal density descriptor, based on which we introduce a dual-network model for correcting the B3LYP functional. Benchmark tests demonstrate that the resulting ML-corrected B3LYP functionals exhibit improved predictive accuracy across a wide range of thermochemical and kinetic energies, particularly in characterizing non-covalent interactions. This work thus emphasizes the importance of extending beyond semilocal density-energy mappings to enhance both predictive power and generalizability of density functional approximations.