Designing forcefields in an age of cheap computing
- John Harding (Sheffield University, United Kingdom)
- Christopher Handley (University of Sheffield, United Kingdom)
- Colin Freeman (University of Sheffield, United Kingdom)
Please visit the conference webpage if you are interested in registering.
If you wish to present a talk or poster then please email Dr Colin Freeman (email@example.com) an abstract and indicate if you would prefer a talk or poster.
Robust, reliable forcefields are central to successful atomistic simulations. This workshop will bring together leading experts to discuss the impact of increasing computer power, both in terms of speed and data storage, on the development, validation and use of forcefields in molecular simulation. Forcefields always require a tradeoff between accuracy and computational cost. When the availability and power of computing rapidly increases, the terms of this tradeoff change. Reviews of forcefields tend to confine attention to classes of materials, for example: soft matter, ceramics, metals and semiconductors, clays, solutions and hetero-systems. Since forcefields always involve a simplified description of the true energy surface, a given forcefield comes with an implicit domain of applicability which is usually discovered by (sometimes bitter) experience. For example, reactive forcefields  are designed to model bond making and breaking - which is beyond the ability of traditional forcefields. However, this new capability is bought at the cost of considerable increase in complexity of the functional form. Some advanced forcefields replace the point charge representation of atoms by three-dimensional electron densities, described using spherical harmonics and obtained by partitioning electron densities between atoms. This approach can also include polarisation effects, but the energy from multipolar electrostatic interactions does not converge at short range.
As forcefields have become more complex, machine learning has been used to optimise transferable parameters for the underlying functions[10,11]. Genetic algorithms are useful for searching the parameter space. Multi-objective optimisation is also valuable since the user can determine the quality of the fitting procedure with respect to structure and energy. Machine learning enables a new route to forcefield design, but requires large amounts of ab initio data to train machine learning methods, and establish the non-linear relationships between atomic positions and energy. This approach has enabled more transferable forcefields, and with growing computer power, the cost of data generation has diminished. However there are still limits to the number of atom types that can be accommodated in a simulation.
The workshop will comprise six sessions, focusing on the problems and opportunities in forcefield design and function to identify common challenges and map the way forward. The traditional justification for forcefields is that they enable simulations to reach the length and timescales needed for adequate sampling of phase space while still offering acceptable accuracy in describing the potential energy surface. Questions to be addressed include;
- Do we need explicit functional forms for forcefields and if so when? Simple functional forms offer fast evaluation of forces but are tied to specific models of crystal binding. For hetero-systems, it is often unclear what functional form is suitable.
- What role can machine learning techniques play in developing new forcefields? Traditionally forcefields have used functional forms chosen a priori with parameters that must be fitted. Machine learning can both enable an automated and efficient route to parameter fitting, and suggest functional forms from ab initio training data.
- How do we design forcefields to take best advantage of new computer architectures such as GPUs? GPUs can generate training data efficiently – this is critical for machine learning approaches. The construction of forcefields should take advantage of the GPU style of partitioning data where possible.
- What information should be used in developing and validating forcefields and how should it be used? Traditionally, empirical data is used. This can limit the configurations and chemical behaviour available for validation. Ab initio simulations give wider choice but limit the forcefield to the physics they contain. Combining the two can lead to consistency problems.
- How desirable is transferability for a forcefields and how do we assess it? We must assume that parameters obtained from the training data-set can be transferred to simulate all configurations encountered in production runs. We often want to use forcefield components for many materials. Can we design forcefields that “degrade gracefully” - that do not produce nonsense when unusual configurations are generated?
- How do we map out the domain of applicability of a forcefield? Since forcefields are imperfect representations of the true potential energy surface, they work well for some applications but not for others. There is much implicit knowledge in research groups. How can this be captured effectively?
- How do we ensure that users have all the information required to use a forcefield? Such information should be reported in the literature but often is not. For example, cut-off procedures may not be fully specified. It may be necessary to use mixing rules to obtain missing parameters and specific methods to generate atomic charges. Sample input files for specific programs are unambiguous (or should be) but are not always provided. There is a role for databases here.
- Can we identify a roadmap for forcefield development? This is the major deliverable for the workshop. Such a roadmap would enable workers to plan a joint strategy for forcefields.
We invite submissions of short talks and posters that address the above topics. Submitted talks will be selected by the organising committee in the event of the program being over subscribed. Submitted talks will each have a 15 minute slot which includes 3 minutes for questions. This should be seen as an opportunity to present novel solutions for atomistic simulations, or outline where the current state of the art models fail for important simulation tasks.
1. Wang J.M. et al.; J. Comp. Chem. 25, 1157 (2004)
2. Gale JD, Phil. Mag. 73 3 (1996); Madden P.A., Wilson M.; Chem. Soc. Rev. 25, 339
3. Plimpton S.J., Thompson A.P.; MRS Bulletin, 37 513 (2012)
4. Cygan, R.T. et al.; J. Phys. Chem. B 108, 1255 (2004).
5. See http://www1.lsbu.ac.uk/water/water_models.html and references therein for a good
discussion of water models.
6. Freeman, C.L. et al.; J Phys Chem C 111 11943 (2007)
7. Liang T. et al.; Ann. Rev. Mater. Sci. 43, 1, (2013).
8. Ponder, J. W., Case, D. A.; Adv. Protein Chem. 66, 27 (2003).
9. Popelier, P. L. A; Int. J. Quantum Chem. 115, 1005 (2015).
10. Handley, C. M., Deeth, R. J. A; J. Chem. Theory Comput. 8, 194 (2012).
11. Larsson, H. R., van Duin, A. C. T. & Hartke, B.; J. Comput. Chem. 34, 2178 (2013).
12. Rice, B. M. et al.; J. Chem. Theory Comput. 11, 392 (2015).
13. Handley, C. M. & Behler, J.; Eur. Phys. J. B 87, 1 (2014).