Abstract:

Quantum mechanics is an accurate theory to compute the physical properties of molecules, but in general is prohibitively expensive when the properties are needed over a large dataset. Instead, molecular mechanics force fields are widely used as an approximation, but developing parameters that give similar accuracy to QM while not overfitting is a major challenge in force field development. Besides determining what the parameter values should be, a potentially more challenging task is determining how many parameters are optimal and where to apply them given an arbitrary molecule. In this thesis, the necessary procedures of automated data-driven force field design are developed, including both efficient chemical space coverage for dataset design and valence parameter selection, initialization, and optimization. Direct computation of force constants for harmonic bonds and angles in addition to dihedral amplitudes and periodicities is shown using a reference QM Hessian. To demonstrate the approach presented in this thesis, a force field was fit to an alkane dataset that started from a minimal number of parameters. The resulting force field achieved a remarkable improvement in accuracy in geometries, forces, and vibrational frequencies compared to an existing force field. Looking forward to extend this work to broader chemical space, a procedure to efficiently generate energy datasets is proposed which utilizes the nudged elastic band method to cover relevant conformational space without the need to scan dihedral angles.