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

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Engineering

Background:

• ReaxFF is a computationally efficient force field for simulating reactive dynamics.
• Optimizing ReaxFF parameters is critical for accuracy but challenging.
• Existing methods like genetic algorithms (GAs) aim to find optimal parameters in a complex search space.

Purpose of the Study:

• To systematically compare the performance of three parameter optimization methods: GA (OGOLEM), Monte-Carlo Force Field Optimizer (MCFF), and Covariance Matrix Adaptation Evolutionary Strategy (CMA-ES).
• To evaluate the reproducibility and convergence behavior of these methods.
• To provide guidelines for selecting and using these optimization techniques.

Main Methods:

• Systematic comparison of OGOLEM (GA), MCFF, and CMA-ES using three literature training sets.
• Repetition of optimizations with varied random seeds and initial parameter guesses to assess reproducibility.
• Analysis of numerical noise impact and strategies for its reduction, such as using unambiguous geometry optimizations.

Main Results:

• No single optimization run should be trusted blindly due to common issues like irreproducibility, poor convergence, or premature convergence.
• GA methods demonstrate the lowest risk of getting trapped in local minima.
• CMA-ES achieved the lowest errors in two-thirds of the cases, though not consistently.
• Numerical noise can be reduced by using unambiguous geometry optimizations.
• Multiple near-optimal parameter sets can be found, suggesting avenues for training set improvement and overfitting detection.

Conclusions:

• CMA-ES and GA offer distinct advantages in ReaxFF parameter optimization, with CMA-ES often reaching lower errors and GA providing more robust convergence.
• Careful consideration of optimization strategies, including reproducibility checks and noise reduction, is essential for reliable ReaxFF parameter development.
• The findings offer practical guidelines for researchers using these methods and highlight the potential for improved training set design and overfitting detection.