Atomistic simulations based on quantum and classical mechanics have increasingly become prominent in the design of materials prior to experimental testing and discovery. Density functional theory (DFT) is the method of choice for simulating periodic structures of metals, alloys, ceramics, semiconductors, insulators, and other types of solid-state materials, and calculating structural, energetic, electronic, magnetic, mechanical, optical, dielectric, and defect properties at quantum mechanical accuracy [1]. DFT is typically performed at the tens to hundreds of atoms scale since it still scales poorly with system size. Crystalline renditions of real-life materials are often approximate, and results depend heavily on the level of theory used, e.g., electronic and optical properties are far more accurate when using nonlocal hybrid functionals as opposed to semi-local functionals [2]. Typical challenges in performing DFT computations arise in the form of (a) solving the Kohn-Sham equations efficiently, especially for large systems, (b) using advanced and very expensive levels of theory for accurate electronic structure calculations, and (c) bridging the gap between idealized theoretical structure-property assessments and real-life experimental conditions.

It is now also routine to overcome the expense of DFT simulations by coupling them with machine learning (ML) and AI models for accelerated prediction, either via rapid differential equation solvers or by direct prediction of charge density, forces and energies, and properties of interest [3-5]. ML force fields trained on enormous quantities of first principles data have been frequently applied to perform accelerated MD simulations [6] or geometry optimization of new structures at a fraction of the cost of full DFT [7]. Models for property prediction at DFT accuracy are widely available for use by non-experts [8]. ML methods have helped bridge different scales of simulations and levels of theory, often using transfer learning or multi-fidelity learning [9]. AI/ML-augmented first principles computations have transformed materials discovery, leading to better performing materials for energy storage and conversion, solid-state catalysis, and other applications related to semiconductors and electronics [10-12].

This workshop aims to bring together several leading researchers working on using AI and ML for accelerating materials simulation and discovery. The time is ripe for discussions on the best methods currently available, how to use and advance them, the challenges that remain, and what the next generation of computational materials science researchers should be focusing on. Speakers and attendees will come from materials science, chemistry, physics, and computer science backgrounds or departments. Part of the focus will be on open-source tools and software for automated and accelerated materials simulations.

Some specific topics of interest include: 

·      Recent developments in crystal graph neural network models.

·      Tools for performing high-throughput ab initio simulations.

·      State-of-the-art in representations or descriptors for materials.

·      State-of-the-art machine learning interatomic potentials.

·      State-of-the-art computational materials databases and their future.

·      Novel fast and accurate DFT functionals developed using ML.

·      Open-source tools for property prediction at DFT accuracy.

·      Combining DFT, ML, and experimental data to predict synthesizability and synthesis routes for new materials.

·      Multi-fidelity learning, transfer learning, and model fine-tuning.

·      The use of LLMs in driving AI-augmented atomistic simulations.