434233 Heterogenous Morphology By Design - Harnessing Cloud to Quantify and Optimize Morphology

Tuesday, November 10, 2015: 10:10 AM
255B (Salt Palace Convention Center)
Olga Wodo, Mechanical and Aerospace Engg Dept, SUNY Buffalo, Buffalo, NY and Baskar Ganapathysubramanian, Mechanical Engg Dept, Iowa State University, Ames, IA

Internal spatial organization of material is known to play a key role in engineered and natural heterogenous systems. Understanding interfacial phenomena and tailoring such heterogenous systems to obtain a desired functionality is a crucial element of energy-related research, advanced manufacturing, micro-fluidic technologies, as well as for personalized medicine.

We present the generic approach to screen large data sets of morphologies. Our approach is based on a MapReduce algorithm. Inherent parallelism of MapReduce and its support for cloud computing makes it a very attractive approach  to perform the efficient and fault tolerant accelerated analysis. We present how to harness it to establish process-structure-property link in efficient and generic way. We illustrate our approach by analyzing structure-property link of organic photovoltaics. In the mapping, stage we emit large set of representative morphologies along (value) with morphology descriptors (key)). We use a graph based strategy to extract physically meaningful morphology descriptors.  In reducing stage, we collect morphologies of the same key and perform full-scale analysis of its properties. This full scale characterization is accomplished using a drift-diffusion model that virtual interrogates the morphology to construct the J-V characteristics.

Subsequently, we use optimization techniques coupled with computational modeling to identify the optimal structures for high efficiency solar cells. In particular, we use adaptive population-based incremental learning method linked to graph-based surrogate model to evaluate properties for given structure.  We study several different criterions and find optimal structure that that improve the performance of currently hypothesized optimal structures by 29%.

Extended Abstract: File Not Uploaded