DEVELOPING A HYDROGEN BOND POTENTIAL USING ARTIFICIAL NEURAL NETWORKS

Cornell University Ithaca, Ithaca NY

Investigators

Linked publications & trials

Abstract

Artificial Neural Networks were introduced to develop a new form for the hydrogen bond potential based on the potential energy surfaces(PES) of water-water and peptide-water systems. The potential energy function has three terms which represent electrostatic interaction, nonbonded interaction and polarization. For the first two terms, classical 1-6-12 pairwise potential energy functions were used. Since the polarization is not an atom pair property, artificial Neural Networks have been applied to describe the complicated potential energy surface instead of analytical pairwise potential energy functions. To calculate the net atomic charges of water, a Potential Derived Charge(PD) method has been applied with some constraints. The charges from the method show good agreement with ab initio electrostatic potentials and also show prominent agreement with experimental dipole and quadrupole moments compared to other methods which have been developed by others. To test the potential energy function, Molecular Dynamics and Monte Carlo simulations will be performed. From the simulations, some physical properties (heat capacity, heat of vaporization and radial distribution function at various temperature) will be calculated and will be compared to experimental data. Hydration energy of organic or small biomolecules can be calculated with this potential and then compared to experimental data. Also minimum-energy structures for the water dimer, trimer and tetramer will be calculated with this potential function and compared to experimental data.