Coarse-Grained Force Field Calibration Based on Multiobjective Bayesian Optimization to Simulate Water Diffusion in Poly-ε-caprolactone

Abstract

Molecular dynamics at the atomistic scale is increasingly being used to predict material properties and speed up the material design and development process. However, the accuracy of molecular dynamics predictions is sensitively dependent on the force fields. In the traditional force field calibration process, a specific property, predicted by the model, is compared with the experimental observation and the force field parameters are adjusted to minimize the difference. This leads to the issue that the calibrated force fields are not generic and robust enough to predict different properties. Here, a new calibration method based on multiobjective Bayesian optimization is developed to speed up the development of molecular dynamics force fields that are capable of predicting multiple properties accurately. This is achieved by reducing the number of simulation runs to generate the Pareto front with an efficient sequential sampling strategy. The methodology is demonstrated by generating a new coarse-grained force field for polycaprolactone, where the force field can predict the mechanical properties and water diffusion in the polymer.

Machine Learning Directed Optimization of Classical Molecular Modeling Force Fields

“…Here, we chose nondominated status and error thresholds for all properties. Alternative strategies include creating a weighted sum of errors in the properties based upon the desired application and domain knowledge, ranking force fields by their error in the various properties studied via statistical tests, 54 evaluating the force field's performance for properties not included in the optimization procedure, or selecting parameter sets based upon a measure of compatibility with the force fields being used for other components of a system. One could also consider chemical intuition when selecting the final parameter sets, e.g., for HFC-125, perhaps a parameter set with more similar values for both fluorine atoms would be preferred.…”

“…In Bayesian optimization, which is a special case of surrogate-assisted optimization, the uncertainty estimates from GPR (or a similar model) are directly used to balance exploration and exploitation. Recent work demonstrates Bayesian optimization can efficiently calibrate force field parameters in coarse-grained models. − Moreover, computationally inexpensive surrogate models can enable multiobjective optimization algorithms that go beyond ad hoc weighting to systematically explore trade-offs when calibrating multiple physical properties.…”

Emerging Era of Biomolecular Membrane Simulations: Automated Physically-Justified Force Field Development and Quality-Evaluated Databanks

“…Automated force field calibration has been done since the late 1990s but has thus far mostly remained in the realm of small molecules; that said, also force fields for complex unstructured biomolecules, such as lipids, have been developed using automated approaches. ,, Gradient-based minimizations, ,,, gradient-free methods such as Simplex − and evolutionary algorithms, ,,,, surrogate or metamodel strategies, ,,, and their various combinations have been used for this purpose (Table ). In addition to the extreme complexity of the underlying optimization problem, one of the main hurdles preventing such automated approaches seems to be the computational cost arising from the fact that the optimization loop requires running MD simulations on (possibly thousands) of force field candidates.…”

Machine Learning Directed Optimization of Classical Molecular Modeling Force Fields

“…In Bayesian optimization, which is a special case of surrogate-assisted optimization, the uncertainty estimates from GPR (or a similar model) are directly used to balance exploration and exploitation. Recent work demonstrates Bayesian optimization can efficiently calibrate force field parameters in coarse-grained models [45][46][47]. Moreover, computationally inexpensive surrogate models can enable multiobjective optimization algorithms that go beyond ad hoc weighting [48] to systematically explore trade-offs when calibrating multiple physical properties.…”

“…Here, we have chosen non-dominated status and error thresholds for all properties. Alternative strategies include weighting the different properties according to domain knowledge, ranking force fields by their error in the various properties studied via statistical tests [46], evaluating the force field's performance for properties not included in the optimization procedure, or selecting parameter sets based upon a measure of compatibility with the force fields being used for other components of a system. Ultimately, selecting the final force field for a given application will be system and application dependent and rely heavily on domain expertise.…”