Efficient optimization of van der Waals parameters from bulk properties

Abstract

Due to the computational cost involved, when developing a force field for new compounds, one often avoids fitting van der Waals (vdW) terms, instead relying on a general force field based on the atom type. Here, we provide a novel approach to efficiently optimize vdW terms, based on both ab initio dimer energies and condensed phase properties. The approach avoids the computational challenges of searching the parameter space by using an extrapolation method to obtain a reliable difference quotient for the parameter derivatives based on the central difference. The derivatives are then used in an active-space optimization method which convergences quadratically. This method is applicable to polarizable and nonpolarizable force fields, although we focus on the parameterization of the AMBER force field. The scaling of the electrostatic potential (ESP) of the compounds is also studied. The algorithm is tested on 12 compounds, reducing the root mean squared error (RMSE) of the density from 0.061 g/cm(3) with GAFF parameters to 0.004 g/cm(3) , and the heat of vaporization from 1.13 to 0.05 kcal/mol. This is done with only four iterations of molecular dynamic runs. Using the optimized vdW parameters, the RMSE of the self-diffusion (Dself ) was reduced from 1.22 × 10(-9) to 0.78 × 10(-9) m(2) s(-1) and the RMSE of the hydration free energies (ΔGsolv ) was reduced from 0.30 to 0.26 kcal/mol. Scaling the ESP to improve dimer energies resulted in the Dself RMSE improving to 0.77× 10(-9) m(2) s(-1) , but the ΔGsolv worsened to 0.33 kcal/mol.