Predicting Selectivity with a Bifurcating Surface: Inaccurate Model or Inaccurate Statistics of Dynamics?

Area of Science:

• Computational Chemistry
• Chemical Dynamics
• Machine Learning Applications

Background:

• Post-transition-state bifurcation (PTSB) reactions present challenges for classical rate theories like transition state theory.
• Quasiclassical trajectory molecular dynamics (QCT-MD) is crucial for understanding complex reaction mechanisms but faces computational cost barriers.
• Direct dynamic simulations struggle with generating statistically significant trajectory data, hindering comparisons with theoretical predictions.

Purpose of the Study:

• To investigate PTSB in Schmidt-Aubé reactions using an enhanced computational approach.
• To overcome the computational limitations of traditional QCT-MD simulations for reactions with complex energy surfaces.
• To improve the accuracy and statistical reliability of predicting reaction selectivity in PTSB systems.

Main Methods:

• Employed machine learning, specifically kernel-ridge regression (KRR), to predict atomic forces.
• Integrated KRR into QCT-MD simulations to accelerate the calculation of atomic forces.
• Significantly increased the number of trajectories simulated to enhance statistical reliability.

Main Results:

• Achieved a >100-fold acceleration in simulating molecular dynamics by using KRR for atomic force calculations.
• Successfully predicted branching ratios for PTSB reactions with greatly reduced statistical errors.
• Demonstrated the KRR-aided QCT-MD approach's effectiveness in enhancing statistical reliability.

Conclusions:

• The KRR-aided QCT-MD method significantly enhances statistical reliability for studying PTSB reactions.
• This computational strategy allows for more confident testing of theoretical models predicting reaction selectivity.
• Dynamical properties influencing branching ratios in PTSB reactions are more clearly elucidated through this advanced simulation technique.