Computational Tools for Nanoscience

Machine Learning with Data Science Methods

Machine learning includes automated learning methods, such as genetic algorithms and neural networks. By combining machine learning with methods from data science such as pattern recognition, computer vision, and multi-objective optimization, we can carry out and analyze efficient atomistic simulations, as well as develop atomistic models from multimodal imaging data.

Reinforcement Learning as well as other optimization approaches such as genetic algorithms are used to develop efficient potential energy surface models or force fields (FFs) for use in molecular dynamics simulations. Inputs for FF models can include both empirical and first-principles information. We successfully developed FFs for novel 2D materials, such as stanene, and a highly efficient and accurate FF for water, allowing mesoscopic mechanisms (e.g., ice-grain formation) to be elucidated.

We have developed a framework for connecting data science/machine learning techniques, multi-modal imaging, and first-principles and atomistic modeling. It allows the integration of multiple experimental characterization results and first-principles calculations via a common data-science-based platform. This platform involves multi-objective genetic algorithms, scale-invariant feature transform, compressive sensing, and dictionary learning. The approach allows, for example, 3-D atomistic structures not explicitly evident in the experimental data to be inferred.

Finally, housed in CNM is a high-performance computing cluster (Carbon) being used for real-time, on-demand data processing.

Tools and Capabilities

• Carbon, High-Performance Computing Cluster (2600 cores, 30 GPUs, ~30 teraflops)
• BLAST, a machine-learning-based toolkit for developing force fields from data sets, including optimization and validation protocols
• FANTASTX, machine learning/artificial intelligence framework to determine atomistic-level structures from multi-modal experimental and theoretical data