Presentation
Facilitating high performance computing for the sciences: a case study in machine learning for materials science
SessionPoster Reception
Event Type
Posters
Reception
Poster
Student Poster
TimeTuesday, July 306:30pm - 8:30pm
LocationCrystal Foyer and Crystal B
DescriptionThe Northeast Cyberteam Initiative is a National Science Foundation funded program aimed at creating a pool of Research Computing Facilitators (RCFs) to help scientists at small and mid-sized universities in the northeast migrate their computational work to high performance facilities in order to more efficiently conduct their research. At the University of Maine a Northeast Cybterteam RCF assisted a computational materials scientist in parallelizing and optimizing their single computer machine learning (ML) algorithms for deployment in a high performance computing (HPC) cluster.
Traditionally, the physical properties of solids are calculated from solving the many-body Schrödinger Equation using their atomic compositions as inputs. However, these first-principles methods are computationally demanding and the results cannot be generalized to predict the properties of novel materials without performing additional time-intensive computations. This research project aimed to develop novel material descriptors and ML models to accurately predict the properties of materials in a fraction of the runtime of traditional methods. The development of such descriptors and models is a key first step in high-throughput discovery of novel energy materials.
This research project used large datasets (~128k molecules/materials) and involved significant computation time. It therefore benefited greatly from being moved to the University of Maine HPC cluster. Through this collaboration the RCF was able to dramatically speed up material descriptor generation and ML model training, which in turn allowed for faster generation and evaluation of descriptors and models. This case study provides an excellent example of how RCFs can help accelerate computationally demanding scientific research.
Traditionally, the physical properties of solids are calculated from solving the many-body Schrödinger Equation using their atomic compositions as inputs. However, these first-principles methods are computationally demanding and the results cannot be generalized to predict the properties of novel materials without performing additional time-intensive computations. This research project aimed to develop novel material descriptors and ML models to accurately predict the properties of materials in a fraction of the runtime of traditional methods. The development of such descriptors and models is a key first step in high-throughput discovery of novel energy materials.
This research project used large datasets (~128k molecules/materials) and involved significant computation time. It therefore benefited greatly from being moved to the University of Maine HPC cluster. Through this collaboration the RCF was able to dramatically speed up material descriptor generation and ML model training, which in turn allowed for faster generation and evaluation of descriptors and models. This case study provides an excellent example of how RCFs can help accelerate computationally demanding scientific research.