深度神经网络结合迁移学习和基因遗传算法加速金属团簇结构全局优化

Accelerated Global Optimization of Metal Cluster Structures via the Deep Neural Network Complemented with Transfer Learning and Genetic Algorithm

在团簇科学中,搜索金属团簇的全局最小能量结构具有重要的意义.最近,深度神经网络结合迁移学习的全局优化方法被开发出来,以提高金属团簇结构的优化效率,该方法可以大幅减少训练深度神经网络所需的样本数量.为了进一步提高势能面的采样效率和深度神经网络结合迁移学习的全局优化方法的全局搜索能力,本文提出了一种将基因遗传算法嵌入到深度神经网络结合迁移学习中的全局优化方法.在Pt_(n)(n=9-15)团簇的全局优化中,该方法只需要深度神经网络结合迁移学习的全局优化方法一半的样本数量就能优化得到全局最优结构,同时节省了约70%~80%的计算成本,这表明全局搜索能力的显著提升。在Pt_(14)团簇势能面上的采样结果显示,该方法搜索到低能量的样本(占比25%)比深度神经网络结合迁移学习方法的样本(占比<1%)更多.在Pt_(16)和Pt_(17)团簇的全局最优结构搜索中,本文报道了文献中尚未报道的新结构,证明本论文建立的方法的可行性和先进性.

Searching the global minimum(GM)structures of metal clusters is of great importance in cluster science.Very recently,the global optimization method based on deep neural network combined with transfer learning(DNN-TL)was developed to improve the efficiency of optimizing the GM structures of metal clusters by greatly reducing the number of samples to train the DNN.Aiming to further enhance the sampling efficiency of the potential energy surface and the global search ability of the DNN-TL method,herein,an advanced global optimization method by embedding genetic algorithm(GA)into the DNN-TL method(DNN-TL-GA)is proposed.In the case of the global optimization of Pt_(n)(n=9–15)clusters,the DNN-TL-GA method requires only a half number of samples at most with respect to the DNN-TL method to find the GM structures.Meanwhile,the DNN-TL-GA method saves about 70%–80%of computational costs,suggesting the significant improved efficiency of global search ability.There are much more samples distributed in the area of the potential energy surface with low energies for DNN-TL-GA(25%for Pt_(14))than for DNN-TL(<1%for Pt_(14)).The success of the DNN-TL-GA method for global optimization is evidenced by finding unprecedented GM structures of Pt_(16) and Pt_(17) clusters.