AIQM2: Better Reaction Simulations with the 2nd Generation of General-Purpose AI-enhanced Quantum Mechanical Method

Reaction simulations provide insight into the mechanisms, guide reaction optimization and catalyst design, and assist in predicting reaction outcomes. Traditional quantum chemical methods for reaction simulation face problems of inferior accuracy and high computational cost. While the rise of the universal machine learning models alleviates the problem of the cost, they mostly remain at the density functional theory level, i.e., typically still far from chemical accuracy. Few of them such as AIQM1 and ANI-1ccx, although targeting coupled cluster level, still reportedly have sub-optimal performance for the reaction barriers. Thus, here we report AIQM2, the general-purpose AI-enhanced Quantum Mechanical Method 2, which utilizes the delta-learning framework to correct the modified GFN2-xTB baseline to the CCSD(T)/CBS level. AIQM2 exhibits overall better performance than its predecessor AIQM1, especially for the description of the transition states and barrier heights, and bypasses common DFT approaches with double-zeta quality basis sets at the cost of semi-empirical methods. The superior speed and accuracy of AIQM2 allow us to propagate a thousand trajectories overnight to revise the product distribution of the bifurcating pericyclic reaction previously investigated by a slower and less accurate B3LYP-D3/6-31G*. AIQM2 is now publicly available in the open-source software MLatom at https://github.com/dralgroup/mlatom and the calculations can be performed online at the https://XACScloud.com.