Quasiclassical Trajectory Simulation as a Protocol to Build Locally Accurate Machine Learning Potentials

Abstract

Direct trajectory calculations have become increasingly popular in recent computational chemistry investigations. However, the exorbitant computational cost of ab initio trajectory calculations usually limits its application in mechanistic explorations. Recently, machine learning-based potential energy surface (ML-PES) provides a powerful strategy to circumvent the heavy computational cost and meanwhile maintain the required accuracy. Despite the appealing potential, constructing a robust ML-PES is still challenging since the training set of the PES should cover a broad enough configuration space. In this work, we demonstrate that when the concerned properties could be collected by the localized sampling of the configuration space, quasiclassical trajectory (QCT) calculations can be invoked to efficiently obtain locally accurate ML-PESs. We prove our concept with two model reactions: methyl migration of i-pentane cation and dimerization of cyclopentadiene. We found that the locally accurate ML-PESs are sufficiently robust for reproducing the static and dynamic features of the reactions, including the time-resolved free energy and entropy changes, and time gaps.

Transforming organic chemistry research paradigms: Moving from manual efforts to the intersection of automation and artificial intelligence

“…However, building a robust ML-PES comes with its own set of hurdles, as the training set for the potential energy surface must sufficiently cover an all-encompassing configuration space. In a landmark breakthrough in 2023, X. Hong and T. Hou charted a pioneering path [90], as shown in Figure 11. They proposed an innovative approach involving the use of quasiclassical trajectory (QCT) calculations when the required properties can be determined by localised sampling of the configuration space.…”

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

“…In studying reaction dynamics, such training sets can sometimes be large. 41 However, a study by Zhang et al demonstrated that the size of the training set can be reduced by limiting the degree of freedom in the reaction space. 41 This can be achieved by running quasiclassical direct dynamics trajectories initialized at TS and proceeding in the direction of forming products.…”

“…41 However, a study by Zhang et al demonstrated that the size of the training set can be reduced by limiting the degree of freedom in the reaction space. 41 This can be achieved by running quasiclassical direct dynamics trajectories initialized at TS and proceeding in the direction of forming products. To the best of our knowledge, an approach using a ML-based force field for reactions with PTSB has not been reported yet.…”

“…To develop an accurate ML model to predict, it is generally believed that the configurational space spanned by the training examples needs to cover that of ML-predicted samples. In studying reaction dynamics, such training sets can sometimes be large . However, a study by Zhang et al.…”