|About this Abstract
||2019 TMS Annual Meeting & Exhibition
||Computational Approaches for Big Data, Artificial Intelligence and Uncertainty Quantification in Computational Materials Science
||A Machine Learning Framework to Improve nanoHUB Prediction Capabilities Using Existing Tool Data
||Saaketh Desai, Sam Reeve, Alejandro Strachan
|On-Site Speaker (Planned)
The cloud computing infrastructure of nanoHUB.org enables online simulation of materials and nano electronic devices across the world. With more than one million simulation jobs across 500+ tools, a significant amount of data has been generated by all the users of the various materials and device simulation tools. This work describes the ability to automatically leverage past user runs to extend the capability of nanoHUB, allowing for predictions by learning correlations and underlying models between inputs and outputs for each tool. Machine learning algorithms such as neural networks and Gaussian processes are trained over existing run data and tested using a 10-fold-cross-validation scheme. The surrogate models created will allow fast results, with uncertainties, if users run simulations using a similar set of inputs to a previous run. Further, this capability can highlight areas of exploration, or lack thereof, with the tool and the sensitivities and limits of these models.
||Planned: Supplemental Proceedings volume