(566j) Exploring the Reproducibility of Molecular Simulations Using the Molecular Simulation Design Framework (MoSDeF)

Molecular simulations are increasingly used to accurately predict the properties of novel materials and understand molecular based phenomena beyond the level of current imaging techniques. Because of this, a wide array of techniques, methods, and software has been published and distributed to the scientific community. Specifically, several molecular dynamics (MD) and Monte Carlo (MC) molecular simulation codes have been made open-source and community developed in order to reduce the workload, overhead, and potential errors with in-house computational codes. Recently, Schappals et al. published work on the comparison of results for density calculations from different force fields with varying groups performing the same calculations using different simulation engines[1]. They found reasonable reproducibility, with a standard deviation of the relative errors (STD-RE) of 0.95%, but the results did not fall within statistical agreement across the engines they studied. They concluded that these systematic errors are inherent to molecular simulations once a certain degree of complexity is reached and stem from a combination of code implementation, software, and human errors. d is a software package designed to standardize inputs to different simulation software, and has the benefits of implementing models in a standardized, well tested, and scriptable way, following the principles of TRUE simulations[2,3]. This work shows that using MoSDeF to initialize a workflow results in consistent engine differences (STD-RE 0.099%) for even more complex arrays of models and methods. Comparing the bias as synonymously as possible to the methods employed by Schappals et. al., MoSDeF gives a 9.7-fold decrease in relative error on a per simulation basis, and a 43-fold decrease in error using best practices replicate averaging. Additionally, the reduction in systematic errors in the model implementation allows us to explore the contributions of other errors from more subtle sources, such as the effect of flexible vs rigid bonds (0.2%-0.8%), the effect of different long range correction treatment (1.2-3.5%), and the effect of slightly altered SPC/E forcefield parameters (0.05%).

References

1. Schappals, M.; Mecklenfeld, A.; Kröger, L.; Botan, V.; Köster, A.; Stephan, S.; Garćıa, E. J.; Rutkai, G.; Raabe, G.; Klein, P.; Leonhard, K.; Glass, C. W.; Lenhard, J.; Vrabec, J.; Hasse, H. Round Robin Study: Molecular Simulation of Thermodynamic Properties from Models with Internal Degrees of Freedom. J. Chem. Theory Comput. 2017, 13, 4270–4280, PMID: 28738147.
2. Summers, A. Z.; Gilmer, J. B.; Iacovella, C. R.; Cummings, P. T.; MCabe, C. MoSDeF, a Python Framework Enabling Large-Scale Computational Screening of Soft Matter: Application to Chemistry-Property Relationships in Lubricating Monolayer Films. J. Chem. Theory Comput. 2020, 16, 1779–1793.
3. Thompson, M. W.; Gilmer, J. B.; Matsumoto, R. A.; Quach, C. D.; Shamaprasad, P.; Yang, A. H.; Iacovella, C. R.; McCabe, C.; Cummings, P. T. Towards molecular simulations that are transparent, reproducible, usable by others, and extensible (TRUE). Mol. Phys. 2020, 118, e1742938.