Accelerating CALYPSO structure prediction with machine learning

The potential of accelerating CALYPSO structure prediction by replacing DFT methods with machine learning was studied. The performance in predicting the potential energy of boron clusters was evaluated with five machine learning methods. Firstly, the original data was represented as structural information metrics with Coulomb matrix. Then the eigenvalue vector pair of the matrix was extracted and used as the input of machine learning algorithms to training model, five algorithms were trained and tested using the same dataset. Also, factors affecting the performance were explored. Finally, a method of comparing the similarity of predicted values and ground truth was proposed based on the characters of potential energy surface (PES), and a confidence model was constructed to validate the best kernel ridge regression (KRR) method. It is suggested that PES fitted by KRR is similar with the PES by DFT, and the confidence of the algorithm is closed to 90% while permissible error is 1 kcal/mol. The result of time test to KRR shows that the method's time complexity is O(n), which is improved by 1 to 2 orders of magnitude compared with DFT methods. © 2021, Jilin University Press. All right reserved.