Vinod Vivin, Lyu Dongyu, Ruth Marcel, R Schreiner Peter, Kleinekathöfer Ulrich, Zaspel Peter
School of Mathematics and Natural Sciences, University of Wuppertal, Wuppertal, Germany.
School of Science, Constructor University, Bremen, Germany.
J Comput Chem. 2025 Mar 5;46(6):e70056. doi: 10.1002/jcc.70056.
Multi-fidelity methods in machine learning (ML) have seen increasing usage for the prediction of quantum chemical properties. These methods, such as -ML and Multifidelity Machine Learning (MFML), have been shown to significantly reduce the computational cost of generating training data. This work implements and analyzes several multi-fidelity methods including -ML and MFML for the prediction of electronic molecular energies at DLPNO-CCSD(T) level, that is, at the level of coupled cluster theory including single and double excitations and perturbative triples corrections. The models for small organic molecules are evaluated not only on the basis of accuracy of prediction, but also on efficiency in terms of the time-cost of generating training data. In addition, the models are evaluated for the prediction of energies for molecules sampled from a public dataset, in particular for atmospherically relevant molecules, isomeric compounds, and highly conjugated complex molecules.
机器学习（ML）中的多保真度方法在预测量子化学性质方面的应用越来越广泛。这些方法，如 -ML 和多保真度机器学习（MFML），已被证明能显著降低生成训练数据的计算成本。这项工作实现并分析了几种多保真度方法，包括 -ML 和 MFML，用于预测 DLPNO-CCSD(T) 水平下的电子分子能量，即包括单双激发和微扰三重校正的耦合簇理论水平。对小分子模型的评估不仅基于预测的准确性，还基于生成训练数据的时间成本方面的效率。此外，还对从公共数据集中采样的分子的能量预测模型进行了评估，特别是对于与大气相关的分子、同分异构体化合物和高度共轭的复杂分子。
