Enhanced Particle Swarm Optimization Algorithm: Efficient Training of ReaxFF Reactive Force Fields.

Area of Science:

• Computational chemistry
• Optimization algorithms
• Materials science

Background:

• Particle Swarm Optimization (PSO) is effective for global optimization but struggles with nonseparable, multimodal functions.
• Existing methods like simulated annealing and genetic algorithms have limitations in complex optimization landscapes.

Purpose of the Study:

• To enhance the rotation-invariant PSO algorithm with isotropic Gaussian mutation operators.
• To improve the search quality and efficiency of PSO on multimodal objective functions.
• To optimize parameters for the ReaxFF-lg reactive force field's dispersion interaction model.

Main Methods:

• Integration of isotropic Gaussian mutation operators into a rotation-invariant PSO framework.
• Benchmarking the enhanced PSO against rotation-invariant PSO, simulated annealing, and sequential one-parameter parabolic interpolation on nonlinear, multimodal functions.
• Application of the optimized algorithm to determine ReaxFF-lg force field parameters using DFT-TS calculations.

Main Results:

• The enhanced PSO algorithm demonstrated superior performance in search quality and efficiency on multimodal benchmark functions.
• Optimized ReaxFF-lg force field parameters accurately described the equations of state for high-energy molecular crystals.
• The improved algorithm outperformed a genetic algorithm in optimizing ReaxFF-lg correction model parameters.

Conclusions:

• Enhancing PSO with isotropic Gaussian mutation significantly improves performance on challenging, multimodal optimization problems.
• The optimized ReaxFF-lg force field provides accurate predictions for high-energy molecular crystals.
• The developed C++ code facilitates the creation and refinement of ReaxFF reactive force fields.