Parameterization of the ReaxFF reactive force field for a proline-catalyzed aldol reaction

Abstract

A parameterization of the ReaxFF reactive FF is performed using a Monte Carlo Simulated Annealing procedure for the modeling of a proline-catalyzed aldol reaction. Emphasis is put on the accurate reproduction of the relative stabilities of several key intermediates of the reaction, as well as, on the description of the reaction path bridging these intermediates based on quantum mechanical calculations. Our training sets include new criteria based on geometry optimizations and short Molecular Dynamics simulations to ensure that the trained ReaxFF potentials adequately predict the structures of all key intermediates. The transferability of the sets of parameters obtained is assessed for various steps of the considered aldol reaction, as well as for different substrates, catalysts, and reagents. This works indeed highlights the challenge of reaching transferable parameters for several reaction steps. © 2016 Wiley Periodicals, Inc.

ReaxFF Parameter Optimization with Monte-Carlo and Evolutionary Algorithms: Guidelines and Insights

“…When simply minimizing the error in eq , there is a significant risk to overfit the parameters. Overfitted parameters have a low error for the training set but still produce nonphysical results in realistic ReaxFF MD simulations. , While we recognize the importance of this problem, this work mainly focuses on testing and understanding strengths and weaknesses of the parameter optimization algorithms. Nevertheless, to avoid overfitting, one may introduce a so-called test set in addition to the training set.…”

JAX-ReaxFF: A Gradient-Based Framework for Fast Optimization of Reactive Force Fields

“…Also, the success of this method is very dependent on the initial guess and the order of the parameters to be optimized. Due to these drawbacks of SOPPI, various global methods such as GAs, − SA, , EAs, and PSO, and search methods based on machine learning − , have been investigated for ReaxFF optimization. For an explanation and evaluation of the most promising of these methods, we refer readers to the prior work by Shchygol et al These methods have been proven to be successful for ReaxFF optimization.…”

“…In both scenarios, the training speed is crucial. As such, development of high-quality and fast optimization methods for ReaxFF has been an active research topic, first starting with SOPPI by van Duin and then continuing with various global optimization methods such as genetic algorithms (GAs), − simulated annealing (SA), , evolutionary algorithms (EAs), and particle swarm optimization (PSO) . More recently, machine learning based search methods have been employed for this purpose. − …”

ReaxFF/AMBER—A Framework for Hybrid Reactive/Nonreactive Force Field Molecular Dynamics Simulations

“…This allowed ReaxFF to be broadly applicable to a wide range of challenging problems. Developed originally for hydrocarbons; the ReaxFF method has been extensively used to investigate complex systems in a wide range of applications including biological systems, − materials, − catalysts, , combustion and batteries …”