Study On Improving Accuracy Of Quantum Chemical Calculations For Non-covalent Interactions Based On Neural Networks

Non-covalent interactions are weak intermolecular interaction that are crucial for bio-molecular structures, super molecules and ligand binding reactions. However, non-covalent interactions are not only experimentally difficult to measure, but also theoretically quite demanding, therefore the advanced ab initio quantum chemical method CCSD(T) and MPn with large basis sets may be required, which cost plenty of machine time and computing resources. To reduce the cost of the calculations so as to perform accurate calculations for non-covalent interactions of large molecular systems, we combine generalized regression neural network GRNN with density functional theory DFT to improve the accuracy of DFT calculations for non-covalent interactions.In this thesis, the calculations of non-covalent interactions are based on P. Hobza’s databases, which include 121 non-covalent binding molecular dimmers. Six DFT methods, M06-2X, B3 LYP, B3LYP-D3, PBE, PBE-D3 and ωB97XD with two small basis sets 6-31G* and 6-31+G* using either water or pentylamine as solvents are set to calculate database molecules. GRNN is used to correct the DFT calculations and obtain results are closer to benchmark interactions calculated by CCSD(T)/CBS method. After correction by GRNN, the root mean square error(RMSE) is significantly reduced, which are decreased at least 68% [M062X/6-31G*(protein)], and the most of reduction is 84%[B3LYP/6-31G*(water)]. The OECD principles are used to evaluate the fitting power, prediction power and stability power of our models. All the validation parameters are larger than 0.9, which means the built models exhibit good stability, robustness and predictivity for non-covalent bond interaction calculations.