What is it about?

It is important to understand the atomic structure of complex molecules. While quantum simulations can help with this, they are expensive and long. Machine learning (ML) can be useful in this case. ML can be used to learn about interatomic potentials, an important property of molecules. This allows for more efficient modeling of atomistic systems. The accuracy of these models is comparable to quantum mechanical methods. However, designing new ML potentials is hard. It requires good knowledge of data science and ML techniques. This is not common in scientists working in chemistry and materials science. This work discusses the best methods to make new artificial neural network (ANN) potentials. New research areas, such as active learning and delta learning, could make big waves for ANN potentials.

Featured Image

Why is it important?

It is critical to help equip scientists with the right ML skills. More knowledge of ANN potentials and applications will speed up their use. This can lead to new exciting research results that would be hard to get with traditional methods. ML interatomic potentials allows accurate simulations of complex systems. It opens new avenues not available before. The available tools and tricks for ML applications is growing. There is a whole collection of free and open source tools and resources. But, ML potentials do not yet exist for many systems. This work gives a recipe for making new ANN potentials. It reviews data and model selection, training and validation, and testing and refinement of models. It also gives practical examples of each of these. KEY TAKEAWAY: Making ANN potentials is a complicated process involving many steps and ML knowledge. This work provides a blueprint towards easier and more automated ANN construction. This would speed up the prediction of materials and molecular properties with an impressive combination of accuracy and speed.

Read the Original

This page is a summary of: Strategies for the construction of machine-learning potentials for accurate and efficient atomic-scale simulations, Machine Learning Science and Technology, July 2021, Institute of Physics Publishing,
DOI: 10.1088/2632-2153/abfd96.
You can read the full text:

Read

Contributors

Be the first to contribute to this page