[gmx-developers] machine-learned potential (GAP)

[Christian Blau]

Hello Fernanda,


For the practical considerations of your approach: The electric field module that is implemented as an IForceProvider 
and gets it's input as IMdModule is a good idea. To see how to get direct access to the array of forces and particle 
positions during a simulation have a look at ElectricField::calculateForces. Once you linked GROMACS against QUIP, 
calling the QUIP force calculation routine with information from ForceProviderInput should work. There is currently no 
infrastructure in place to select close-by water molecules to a given metal ion, which will become a challenge if you 
want to run your code on multiple nodes, however, implementing your method for yourself to run on a local workstation 
appears possible with the current GROMACS version.

Best,
Christian

On 2019-03-07 11:00, Fernanda Duarte wrote:
> Dear GROMACS developers,
> 
> 
> Would it be possible to utilise an additive potential alongside that calculated in GROMACS from standard force-fields? 
> Specifically, we are looking at using a machine-learned potential (GAP) to introduce the effect of a metal 
> ion centre into an MM (water) system. We’d like to train the potential on the difference between an QM/MM calculations 
> with and without Pd, then sum the difference in force experienced by the surrounding atoms to the MM force calculated 
> using the normal MM strategy (and likewise with the energy). In practice: include a standard dummy metal centre and then 
> leave it with 0 vdw/charges and introduce the potential on the top of the empirical one. it will replace the metal + add 
> extra forces to the water close.
> The potential is implemented within QUIP (https://github.com/libAtoms/QUIP), written largely in fortran95. It requires 
> the number of atoms (int), atomic numbers (int, Nx1 array), positions (dp, Nx3 array). We would like to have this as an 
> external source of forces on the central atom and wonder ifthe IForceProvider interface would provide a way to do so. If 
> so, is this possible in the current implementation?
> 
> 
> Many thanks in advance for your help.
> 
> 
> Best wishes,
> 
> Fernanda
> 
> 
> Dr Fernanda Duarte
> 
> Associate Professor of Chemistry
> 
> Fellow and Tutor at Hertford College
> 
> Chemistry Research Laboratory
> Mansfield Road
> University of Oxford
> Oxford OX1 3TA UK
> 
> E:fernanda.duartegonzalez at chem.ox.ac.uk
> 
> T: 01865282770
> 
> W:http://fduartegroup.org
> 
> 
>