Machine Learning for Atomistic Simulation, Materials Design and Autonomous Discovery
This community develops computational tools that use machine learning to simulate atomic behavior, predict material properties, and automate the discovery of new chemical compounds.
The work centers on creating "interatomic potentials" and "force fields" that allow for high-fidelity molecular dynamics simulations without the computational cost of quantum mechanics. Recurring methods include "neural networks," "graph neural" architectures, and "active learning" to optimize these models. Applications span "catalyst design," "crystal structure" prediction, and "inverse design" of materials. The community also integrates "large language" models and "language models" to extract data from scientific literature and drive "autonomous" laboratory workflows. Specific chemical systems frequently addressed include "transition metal" compounds and "reactive molecular" interactions, with a strong emphasis on "density functional" theory as a training data source.
The community’s output is distributed across a wide range of elements, with the largest shares of research found in carbon, hydrogen, and silicon. Carbon research accounts for the highest number of papers within this group, followed by hydrogen and silicon.
The community comprises 4,066 papers, published primarily in npj Computational Materials, ChemRxiv, and arXiv (Cornell University).
Recent work focuses on developing general-purpose "neural network potentials" for diverse material classes, including energetic materials and advanced structural compounds. Newer studies also explore "AI agents" for autonomous discovery and "physics-informed neural networks" to bridge quantum accuracy with computational efficiency in energy materials and electrocatalysis.