“…In the previous calculations, , the potential energies of TBMD were approximated by a tight-binding Hamiltonian, and ReaxFF PES was parameterized and fitted by a database of pre-computed DFT energies, both of which may introduce an error of roughly 0.2 eV diverged from DFT values. The AIMD approach computes the motion of all atoms and potential energies on the fly and enables a rather accurate description of both interaction and energy exchange between the impinging molecule and surface. , However, AIMD requires a large number of DFT calculations, making it computationally expensive, and only a small number of AIMD trajectories is affordable, making a good statistical sampling difficult. , Recently, neural network (NN) methods have been widely used for fitting high-dimensional ab initio PES because of flexible nonlinear analytical functions which can represent a multidimensional data set with high fidelity. − The permutation invariant polynomial-NN (PIP-NN) approach incorporates both the permutation symmetry of the molecule and surface periodicity in the NN method that has achieved great success in representing high-dimensional PESs for interactions between various molecules and surfaces. − To the best of our knowledge, there has been no global DFT PES reported for the O 2 /Pt(111) system, which is undoubtedly desirable for a better understanding of the sticking mechanisms. Here, we present the first six-dimensional PES constructed with thousands of DFT energies based on the Born–Oppenheimer and static surface (BOSS) approximation .…”