Optimization of transferable force fields based on reduced order and surrogate models
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Stuttgart : Universität Stuttgart, Institut für Technische Thermodynamik und Thermische Verfahrenstechnik
Abstract
Klassische transferierbare Kraftfelder werden verwendet, um Vorhersagen über physikalische Eigenschaften und mikroskopische Prozesse sowohl von reinen Substanzen als auch von Gemischen zu treffen. Bei der Entwicklung dieser Kraftfelder werden die nicht gebundenen Wechselwirkungen durch mehrdimensionale Optimierungsverfahren optimiert, wobei rechenintensive Monte-Carlo- und Molekulardynamiksimulationen zum Einsatz kommen. Aufgrund der hohen Korrelation der Kraftfeldparameter kann dieser Optimierungsprozess eine Herausforderung darstellen. Um die Effizienz zu verbessern, verwenden wir physikalische Ersatzmodelle (Surrogate Models), um den Optimierungsprozess zu beschleunigen. Insbesondere wurden neue Modelle für Transporteigenschaften entwickelt, die auf Entropieskalierung basieren. Dies ermöglicht die Berücksichtigung von Transporteigenschaften wie Viskosität, Wärmeleitfähigkeit und Selbstdiffusionskoeffizient bei der Optimierung von Kraftfeldern gleichberechtigt neben statischen Eigenschaften, was bisher nicht möglich war. Darüber hinaus werden Machine Learning Modelle für Transporteigenschaften entwickelt, die auf experimentellen Daten und Simulationsergebnissen basieren. Dies wird es uns ermöglichen, unsere Vorhersagefähigkeiten zu verbessern und Kraftfelder effektiver zu optimieren. Die Modelle basieren auf linearen Multifidelity Gauß-Prozessen, die im reduzierten residuellen Entropieraum verwendet werden.
Classical transferable force fields are utilized to make predictions about physical properties and microscopic processes in both pure substances and mixtures. The process of developing these force fields involves optimizing the nonbonded interactions through multidimensional optimization techniques, utilizing computationally intensive Monte Carlo and molecular dynamics simulations. Due to the high correlation of force field parameters, this optimization process can be challenging. To improve efficiency, physically-based surrogate models were utilized to expedite the optimization process. In particular, new reduced order models for transport properties based on entropy scaling were developed. Now transport properties such as viscosity, thermal conductivity and self-diffusion coefficients can be represented in the optimization of force fields through surrogate models alongside static properties, which was previously not possible. Furthermore, machine-learned models for transport properties based on linear multifidelity Gaussian processes utilized in reduced residual entropy space were developed. Here, experimental data is combined with simulation results which enables to enhance predictive capabilities and optimize force fields more effectively.
Classical transferable force fields are utilized to make predictions about physical properties and microscopic processes in both pure substances and mixtures. The process of developing these force fields involves optimizing the nonbonded interactions through multidimensional optimization techniques, utilizing computationally intensive Monte Carlo and molecular dynamics simulations. Due to the high correlation of force field parameters, this optimization process can be challenging. To improve efficiency, physically-based surrogate models were utilized to expedite the optimization process. In particular, new reduced order models for transport properties based on entropy scaling were developed. Now transport properties such as viscosity, thermal conductivity and self-diffusion coefficients can be represented in the optimization of force fields through surrogate models alongside static properties, which was previously not possible. Furthermore, machine-learned models for transport properties based on linear multifidelity Gaussian processes utilized in reduced residual entropy space were developed. Here, experimental data is combined with simulation results which enables to enhance predictive capabilities and optimize force fields more effectively.