Mitigating error cancellation in density functional approximations via machine learning correction

The integration of machine learning (ML) with density functional theory has emerged as a promising strategy to enhance the accuracy of density functional methods. While practical implementations of density functional approximations (DFAs) often exploit error cancellation between chemical species to achieve high accuracy in thermochemical and kinetic energy predictions, this approach is inherently system-dependent, which severely limits the transferability of DFAs. To address this challenge, we develop a novel ML-based correction to the widely used B3LYP functional, directly targeting its deviations from the exact exchange-correlation functional. By utilizing highly accurate absolute energies as exclusive reference data, our approach eliminates the reliance on error cancellation. To optimize the ML model, we attribute errors to real-space pointwise contributions and design a double-cycle protocol that incorporates self-consistent-field calculations into the training workflow. Numerical tests demonstrate that the ML model, trained solely on absolute energies, improves the accuracy of calculated relative energies, demonstrating that robust DFAs can be constructed without resorting to error cancellation. Comprehensive benchmarks further show that our ML-corrected B3LYP functional significantly outperforms the original B3LYP across diverse thermochemical and kinetic energy calculations, offering a versatile and superior alternative for practical applications.