Extension of Composite Method and Machine-Learned Electron Correlation Model to Fourth-Period Elements

Accurate and efficient correlation energy calculation is a major challenge in quantum chemistry. We propose an extended machine-learned electron correlation (ML-EC) model that estimates CCSD(T)/CBS correlation energy using descriptors from Hartree-Fock (HF) calculations with double-zeta basis sets. While the previous ML-EC model was limited to third-period elements, we extend it to fourth-period elements by modifying the composite method parameters. The optimized parameters accurately reproduce CCSD(T)/CBS correlation energies and correlation energy densities. Trained on G3/05 dataset molecules, the ML-EC model accurately predicts CCSD(T)/CBS correlation energies for test molecules. Reaction energies computed with the ML-EC model surpass DFT methods in accuracy. Additionally, the ML-EC model significantly reduces computational cost, achieving a speedup of over 50 times compared to conventional CCSD(T)/CBS calculations. These results demonstrate that the extended ML-EC model is a reliable and efficient method for correlation energy calculations, particularly for systems containing heavy elements.