Machine Learning Potentials and Large Language Models for Chemical Discovery
This community develops computational tools that use machine learning to predict atomic interactions, simulate chemical reactions, and automate the design of new materials.
The work centers on creating interatomic potentials and force fields that allow for accurate molecular dynamics simulations of complex systems. A significant portion of the research focuses on training neural networks, including graph neural networks and deep learning models, to predict crystal structures and chemical properties. Concurrently, the community applies large language models to extract data from scientific literature, generate research hypotheses, and facilitate autonomous laboratory workflows. These methods are applied to specific challenges such as modeling amorphous silicon, transition metal catalysis, and the discovery of inorganic materials. The integration of active learning and Bayesian optimization enables the efficient exploration of chemical space, reducing the computational cost of traditional density functional theory calculations.
The community's output is most heavily concentrated in silicon research, accounting for 20% of all silicon studies tracked, and germanium research, where it represents 20% of that element's literature. Silicon also contributes the highest absolute number of papers to this group, with 820 entries, followed by carbon with 523 papers.
The community comprises 6,489 papers, with the highest publication volumes appearing in arXiv, npj Computational Materials, and ChemRxiv.
Recent work focuses on developing universal machine learning interatomic potentials for large-scale atomistic simulations and using foundation models to predict thermal conductivity. Other recent studies explore the use of knowledge graphs and large language models to generate scientific ideas and evaluate AI's ability to perform expert-level scientific tasks.