Prediction of Methanol Crystal Structures using Steady State Genetic Algorithms and Neural Networks in the Open Science Grid.

A. M. LUND; J. C. FACELLI; G. I. PAGOLA; M. B. FERRARO; A. M. ORENDT

Ciudad de Buenos Aires

Conferencia; Stat Phys 27 International Conference on Statistical Physics; 2019

We present the preliminary results of using a combination of artificial neural networks (ANN) and Evolutionary Algorithms (EA) to improve the performance of our Modified Genetic Algorithm for Crystals (MGAC) to predict the crystal structures of ethanol. The current version of MGAC [1] has been enhanced and modernized for use with the density functional theory (DFT) software Quantum Espresso [2], but its GA implementation suffers from poor performance in large cloud systems do the synchronization required at every generation. Here we repot how the GA has been reformulated using a variable population size strategy leading to the steady state genetic algorithm (SSGA), which allows for much better scalability and use of computation resources in large scientific clouds like the Open Science Grid (OSG). [3] The SSGA introduces substantial modifications to the GAs in which the traditional concept of generations is eliminated allowing computationally more expensive calculations to be distributed in a non-sequential manner and substantially accelerating the calculations. This new algorithm is being implemented to use the resources of the OSG [3], which allows us to perform a larger volume of local optimizations of candidate crystals than those that were previously made by MAGC. Moreover, to accelerate the convergence we employ neural networks using the SSGA to search for the design the structure and training of the network. The training data consists of the parameters (GA gens) of a set of optimized crystal candidates and employ those parameters before and after the optimization of each of those crystals to use them as input and output of the neural network to train it. The trained network is therefore used to pre-optimize the new candidate crystals that are constantly generated by the MGAC.We present here preliminary results using methanol as a molecular test system of bibliographic interest [4], which is small and has several polymorphs. We expect to find those polymorphs in the right energy order.KEYWORDS. AEE, GA, CRYSTAL PREDICTIONS[1] A.M. Lund., A. M. Orendt, G.I. Pagola, M.B. Ferraro, J.C. Facelli, Crys.Growth.Design,2013,13,2181[2] www.quantum-espresso.org[3] https://opensciencegrid.org/[4] Tzu-Jen Lin, Cheng-Rong Hsing, Ching-Ming Wei, Jer-Lai Kuo Phys.Chem.Chem.Phys., 2016,18,2736