Abstract:The prediction of new compounds via crystal structure prediction may transform how the materials chemistry community discovers new compounds. In the prediction of inorganic crystal structures there are two distinct classes of prediction; Performing crystal structure prediction via heuristic algorithms, using a range of established crystal structure prediction codes, and an emerging community using generative machine learning models to predict crystal structures directly. In this work, we demonstrate the use of a generative machine learning model to produce the starting population of crystal structures for a heuristic algorithm and discuss the benefits, demonstrating the method on eight known compounds with reported crystal structures and three hypothetical compounds. We show that the integration of machine learning structure generation to heuristic structure prediction results in both faster compute times per structure and leads to lower energies. This work provides to the community a set of eleven compounds with varying chemistry and complexity which can be used as a benchmark for new crystal structure prediction methods as they emerge.

What problem does this paper attempt to address?

The paper attempts to address the problem of accelerating the discovery of new compounds in inorganic materials chemistry through crystal structure prediction (CSP). Specifically, the paper explores the method of combining generative machine learning models with the heuristic crystal structure prediction code FUSE to improve prediction efficiency and accuracy. ### Main Objectives of the Paper: 1. **Improve Computational Efficiency**: By using machine learning-generated initial crystal structures instead of traditional random structure generation methods, the computational time required for each structure optimization is reduced. 2. **Lower Energy States**: Machine learning-generated structures are closer to real crystal structures, thus allowing the discovery of lower energy global minimum structures within similar computational times. ### Method Overview: - **Machine Learning Model**: A graph neural network (GNN) model is trained to generate initial crystal structures from two benchmark databases (OQMD and MatB). - **FUSE Code Improvement**: The FUSE code is modified to accept machine learning-generated structures and decompose them into submodules for further optimization. - **Experimental Design**: Experimental tests were conducted on eight known compounds and three hypothetical compounds, comparing the methods of using machine learning-generated structures and traditional random-generated structures. ### Main Conclusions: - In most experiments, using machine learning-generated structures significantly improved computational efficiency, with an average speed increase of 2.2 times. - For six known compounds and all three hypothetical compounds, the method using machine learning-generated structures performed better during optimization. - In some cases (e.g., CoAs2), the speed increase reached 8.3 times. Through these methods, the paper demonstrates the advantages of combining machine learning with heuristic algorithms and provides new ideas and technical means for the efficient discovery of new materials.

Integration of generative machine learning with the heuristic crystal structure prediction code FUSE

Integration of generative machine learning with the heuristic crystal structure prediction code FUSE

A machine learning potential-based generative algorithm for on-lattice crystal structure prediction

Enhancing crystal structure prediction by combining computational and experimental data via graph networks

Exploration of crystal chemical space using text-guided generative artificial intelligence

Crystal structure prediction by combining graph network and optimization algorithm

Self-Supervised Generative Models for Crystal Structures

Machine learning approaches for feature engineering of the crystal structure: Application to the prediction of the formation energy of cubic compounds

3-D Inorganic Crystal Structure Generation and Property Prediction via Representation Learning

Machine-Learned Fragment-Based Energies for Crystal Structure Prediction

A Machine-Learning-Assisted Crystalline Structure Prediction Framework To Accelerate Materials Discovery

Data-Driven Score-Based Models for Generating Stable Structures with Adaptive Crystal Cells

Predicting Synthesizability using Machine Learning on Databases of Existing Inorganic Materials

A Hybrid Machine Learning Approach for Structure Stability Prediction in Molecular Co-crystal Screenings

Structure prediction and materials design with generative neural networks

Crystal structure prediction with machine learning-based element substitution

Accelerated Organic Crystal Structure Prediction with Genetic Algorithms and Machine Learning

Unleashing the power of novel conditional generative approaches for new materials discovery

Generative Hierarchical Materials Search

Adaptive Genetic Algorithm for Crystal Structure Prediction

Geometric Deep Learning for Molecular Crystal Structure Prediction