Embedded Machine Learned Interatomic Potentials
Neural network potentials (NNPs) are neural networks that are trained using machine-learning to reproduce the potential energy of atomic coordinates. Neural network potentials, like the ANI-1ccX model, provide the same accuracy of high-level ab initio CCSD(T) calculations but at one billionth of the cost. We have developed to use a NNP to describe a component of the system but the rest of the system is described using a conventional of molecular mechanical model. We have implemented this through the QM/MM features of NAMD.
Long Range Forces in Machine Learned Potentials
The description of long-range forces has been a challenge in machine learned interatomic potentials, which often limit interactions to a short cutoff distance around an atom. Our group has developed the MLXDM method to treat long-range dispersion interactions within the framework of ANI neural network potentials and general-purpose implementations of the CENT ML QEq model.
https://github.com/RowleyGroup/MLXDM
https://github.com/zeldery/combinenet
Generalized Langevin Methods
Calculating position-dependent diffusion coefficients is a challenge for molecular simulation methods. We have implemented a simple method to compute this property by performing a simulation where the solute is restrained to a position. The diffusion coefficient can be estimated by analyzing the time series of that simulation based on the Generalized Langevin Equation. This is implemented in our code, ACFCalculator.
GitHub
Although this strategy is general, we present the theory and practical considerations in our paper where this method is used to calculation the rates of diffusion of solutes through membranes.
Multiscale Methods
Our group develops the QM/MM interface between CHARMM and TURBOMOLE. Researchers who are interested in using CHARMM–TURBOMOLE should obtain licenses for CHARMM and TURBOMOLE. The QM/MM interface is distributed as part CHARMM. Instructions for compiling and using CHARMM–TURBOMOLE are available here. The capabilities of CHARMM–TURBOMOLE are described in this paper:
Riahi, S., Rowley C.N. The CHARMM–TURBOMOLE Interface for Efficient and Accurate QM/MM Molecular Dynamics, Free Energies, and Excited State Properties. J. Comput. Chem. 2014, DOI: 10.1002/jcc.23716 [PDF preprint]
Researchers who use CHARMM–TURBOMOLE should cite this paper in their work.
Example input scripts and data files for CHARMM–TURBOMOLE calculations can be downloaded from our GitHub repository.
