چكيده به لاتين
Deep learning caused a huge revolution in many aspects of science, image, text, and speech recognition. Computational sciences are not an exception in this revolution. In this research, we demonstrate how a deep learning model is trained using quantum mechanical and especially DFT calculation data. This neural network can learn an accurate and transferable potential for many systems, especially organic molecules. Here the ANI model is introduced and working with this model is described. ANI is a new method designed to develop transferable potentials neural networks. This model uses a modified version of the Behler and Parrinello symmetry functions to construct single-atom atomic environment vectors as a molecular representation. These atomic environmental vectors enable us to train neural networks that include both configurational and conformational space. The result of ANI theory is the creation of potentials called ANI-1, which is trained from a subset of GDB databases with up to 8 heavy atoms to predict total energy for molecules containing four types of atoms: H, C, N, and O. According to studies, the ANI-1 model is chemically accurate compared to quantum computations of reference density functional theory in much larger molecular systems than those presented in the training data set. Then, with the help of this model, the structure of n-hexane is studied in detail and compared with the output of quantum computing. Energy and frequency values were evaluated and compared with the output of quantum computing to evaluate the model. In the following, the potential energy levels for a number of important torsions in n-hexane, 2-methylhexane and 3-methylhexane structures are scanned and a new solution is proposed to improve and develop the classical force fields.
