Force field parameterization

Automated generation of molecular interaction parameters by gradient and cognitive methods

Our research

Within the field of molecular simulation, the term “force field” refers to the mathematical model used to represent molecular interactions. This is true for atomic (all-atom) molecular dynamics and for coarse-grain (beaded) simulations in which groups of atoms are defined as a bead.

All-atom force fields provide atomic resolution and are helpful in elucidating molecular phenomena. However, the force fields can be fairly complex and too computationally expensive to simulate large systems. Coarse-grained force fields generally use a simplified interaction potential and have fewer interaction sites, as the atoms are amalgamated into beads. This reduces the computational expense and enables larger systems to be simulated at the expense of some of the finer detailed physics, which occurs at the atomic resolution.

We utilise both all-atom and a version of coarse graining called dissipative particle dynamics (DPD) force fields to perform molecular simulations. All-atom simulations have been used extensively across multiple fields for several decades. As a result, there are several force fields with good parameters, which have been tried and tested for a variety of systems. However, DPD is a relatively new method without standard force fields. We have therefore devised methods to parameterize a DPD force field automatically for a given application. Our methods optimize DPD force field parameters in reference to a user’s experimental data.

This work is being successfully applied to DPD simulations of micelle formation and partition coefficient (logP) predictions.

Two complementary methods

Cognitive parameterization

Using the global optimization method of Bayesian optimization, we can quickly and efficiently sample points in our parameter space to find an optimal region for a given experimental observable. This method also enables a prior to be specified, enabling expert knowledge to be easily accounted for in the optimization process.

Force balance parameterization

Using local gradient based optimization methods, such as steepest descent and BFGS (Broyden–Fletcher–Goldfarb–Shanno), we can move towards a well-defined minimum on the parameter space’s hyper-surface. The method optimizes the parameters of the DPD force field to a local minimum defined by the parameters hyper-surface.


[1] R.L. Anderson, D.J. Bray, A.S. Ferrante, M.G. Noro, I.P. Stott, P.B. Warren,
Dissipative particle dynamics: Systematic parametrization using water-octanol partition coefficients,”
The Journal of Chemical Physics 147(9), 094503, 2017.

[2] M.A. Johnston, W.C. Swope, K.E. Jordan, P.B. Warren, M.G. Noro, D.J. Bray, R.L. Anderson,
Toward a Standard Protocol for Micelle Simulation,”
The Journal of Physical Chemistry B 120(26), 6337-6351, 2016.

Ask the experts

James Mcdonagh

James Mcdonagh

Alexander Harrison

Alexander Harrison

Breanndan O Conchuir

Breanndán Ó Conchúir

Edward Pyzer-Knapp

Edward Pyzer-Knapp

William Swope

William Swope
IBM Research – Almaden, USA

Michael Johnston

Michael Johnston
IBM Research – Dublin, Ireland