Tuesday, 1 August 2017

Next Generation Force Fields - Designing force fields in an age of cheap computing - Perspective

This past week at Halifax Hall, the University of Sheffield was host to a workshop of experts and world leaders on computational chemical simulations. The aim of the workshop was to present the advances in this field, for the atomistic simulation of bulk materials, surfaces, interfaces, solutions, biomolecules, and more. It was also a chance for the delegates to discuss the next challenges facing the discipline given the increasing power of computers.

Chaired by Dr Colin Freeman, Dr Chris Handley, and Prof. John Harding, the invited speakers included Prof. Nohad Gresh, Prof. Jorg Behler, Prof. Bernd Hartke, Prof. Stefan Goedecker, Dr Peter Brommer, Prof. Paul Popelier, Dr Paul Richmond, and Dr David Mobley.

From the talks and contributions it was clear there are common challenges that all researchers face in the coming years. In terms of the technologies available, the push to GPU computing, driven by machine learning, means that the software used for chemical simulations must be reassessed if they are going to exploit the potential computer power on offer.

Another key issue is how machine learning is used to design new chemical simulations. Machine learning, such as neural networks, or Gaussian process regression, can be trained to extract the underlying non-linear relationship between atomic positions and some property of the system (we are often looking for the relative energy differences for a given configuration of atoms, and the associated forces as these are used to drive the molecular dynamics simulation). The machine learning methods use for training data quantum mechanical simulations. The danger however, is that while the non-linear model is discovered, we retain no information about the true physics at play, which is critical if we wish to have a deeper understanding of the interactions at play within a simulation.

Physics driven models are the alternative, but these models rely on knowing the proper functional forms for describing the interactions between atoms. Classical force fields often use functional forms that have been used over the decades that were initially chosen for computational convenience. Going forward our choice of functional forms used should be reassessed given that computational power is no longer an issue.
The effect of d-electrons on atomic configurations is an
electronic orbital effect and one that is not captured in
traditional force fields.

Related to the two previous points, is one of transferability. Force fields are often fitted, and thus simulate well, a particular chemical system in certain conditions. Knowing the limits of a particular model, and also how badly it will perform when if such conditions are met, is knowledge not often provided when a force field is published. Care must be taken that a model is capable of representing all the relevant physics that describes the system. There may be underlying physics not accounted for by the model explicitly, which hinders the transferability of the model when applied to similar chemical systems.

Transferability, data driven models, and physical driven models, ultimately are tied to how we partition the energy of a force field. By this we mean, how do we cut up the force field into different interactions, and how we even define these interactions. How do we define bonds? Can we accommodate reactivity into our models? What of electrostatic interactions? Should we use multipolar descriptions of charge?

Fitting and automation can be reliant on "wizardry". By that we mean, "to use a force field, and to design derivatives of it, how much highly specific expertise do we require - is the force field only really usable by those who designed it?" This is not ideal, as this slows the development of new models, and also hinders insight into why models work and fail for particular simulations.

Finally there is the topic of the reference data. More often than not, force fields are fitted to quantum mechanically generated reference data. We assume that this data is the "truth". Though, while many of the programs that perform quantum mechanical calculations have over time become closer in agreement with each other, they all share the same issue of accuracy. The fitting of force fields is thus a cyclic issue, where a reassessment of the training should be performed periodically.

Going forward the issues raised during the workshop will help inform CECAM and EU funding on future force field development and how it plays a critical role in computational chemistry, even in an age of high performance computing.

As organisers of the workshop, Colin, Chris and John would like to thank CECAM and CCP5 for the generous funding for the event, the invited speakers for their informative presentations, and the attendees whom we hope will have taken away new ideas and new future collaborations.

Conference Attendees