(582d) Strategies for Efficiently Exploring Structure-Property Spaces Via High-Throughput Screening of Hypothetical Materials Conference: AIChE Annual MeetingYear: 2015Proceeding: 2015 AIChE Annual MeetingGroup: Computational Molecular Science and Engineering ForumSession: Data Mining and Machine Learning in Molecular Sciences II Time: Wednesday, November 11, 2015 - 4:15pm-4:30pm Authors: Kaija, A., University of Pittsburgh Wilmer, C. E., University of Pittsburgh Exploring structure-property spaces is an integral, but challenging, step in application-driven material design. The practically infinite number of possible materials makes it difficult to map a meaningful fraction of the structure-property space. Experimental screening of real materials consumes significant time and physical resources, and although computational screening of hypothetical materials allows one to map a larger space in less time, the number of possibilities is still too large for a brute force approach in either case. In this work, we describe an efficient method for mapping the physical limits of a structure-property space by randomly-generating an initial population of hypothetical materials, simulating their properties, and then generating subsequent populations by mutating those materials whose properties are rare. Our initial application area to test the method focused on methane adsorption in porous media. We considered our finite exploration of the structure-property space sufficiently representative of the infinite space, when adsorption data varied smoothly as a function of structure (e.g., void fraction, pore size, volumetric surface area) everywhere in the structure-property domain (except at the boundaries). Initial populations are generated by populating unit cells with randomly positioned “pseudo” atoms, whose Lennard-Jones force field parameters are also randomly chosen. Methane adsorption at 35 bar and 298 K is simulated for each material using classical grand canonical Monte Carlo simulations. In addition, structure properties such as void fraction and surface area are also calculated for each material. Following an analysis of the initial population, we identified structure-property combinations that were rare, and then mutated materials with those properties to generate new populations. The mutation script perturbs all of a material’s defining parameters (atom site positions, Lennard-Jones parameters, etc.) to a degree dictated by a strength parameter. Once the structure-property space has been “filled,” which we define as the point where every structure-property combination is equally represented, we stop generating new materials (i.e., the method has converged). In this work, we investigate the impact of adjusting the mutation strength parameter on the speed of convergence. The method developed in this work provides an efficient way to map an entire structure-property space using only a small, finite, number of hypothetical materials. This work will help guide the design of materials for methane storage and potentially other gas-related applications in the future.