It has been reported by the latest paper in Phys. Rev. B about a new type of strategy for calculating potential-energy surface. The nerual network, a mathematical model in the field of artificial intelligence has been developed to present a ab-initio standard precision of calculation about potential energy. What's more, this pure mathematical model will take much less time than the regular first-principle calculation. This blog could be regarded as an brief introduction, details could be found in the paper "Phys. Rev. B, 85, 045439 (2012)".
The structure of neural network(NN) applied is illustrated below:
The coordinate of atoms are transformer into the defined symmetric function, which represent the distribution of electronics in atom. These function will be defined beforehand for a complex system, then a combination of DFT calculation and NN will be used to generate training data. 1890188 pieces of information has been generated and applied in the training of NN, which ensure that the NN will cover all kinds of atomic environment.
However, there are still significant drawbacks in this strategy. In the first place, a large number of calculations should be carried out as the training date, so as to guarantee the precision. Further, the precision of the NN system in extremely large and complex system is still unknown. Moreover, the symmetric function should be redefined in a multicomponent system, which will not save enough time compared with DFT calculations. Only when they are adapted in the molecular dynamics and Monte Carlo simulation, the advantage will be showed as a result. For instance the examples of calculation about Cu system is showed below: