Research And Applications Of Material Property Prediction Based On Graph Neural Networks

The development of materials science is an important cornerstone of scientific and technological innovation,and the design and development of high-performance new materials typically require significant human and material resources.Computational simulation methods can provide theoretical guidance for researchers and accumulate a large amount of data resources.The development of machine learning techniques enables people to fully exploit the value of these theoretical and experimental data and is expected to further accelerate the research progress of materials science.Traditional machine learning models rely on manually designed descriptors and require researchers’ expertise to select material features.The model performance depends on the appropriateness of feature engineering and often only applies to certain specific tasks.In this work,we first developed a material property prediction framework based on graph neural networks,which can achieve end-to-end property prediction for both molecules and crystals.The framework uses the message passing mechanism,combined with dynamic attention,a unified embedding layer and other techniques,allowing the model to automatically optimize the relationships between input vectors and dynamically adjust the weight coefficients between atoms.Without relying on elemental physical and chemical properties,it can achieve excellent prediction performance.Secondly,in the application of various property predictions of perovskite materials,we compared the accuracy advantage and generalization ability of the framework with various classical machine learning algorithms.Finally,the framework combined with density functional theory calculations and transfer learning successfully applied to enthalpy prediction of different phases of ruthenium dioxide materials,thereby accelerating the search for new efficient oxygen evolution reaction catalyst materials.This research provides a universal and feasible approach for material property prediction,achieving good results,and explores solutions to limitations such as insufficient data in practical applications.The developed framework and pre-training models have important significance for accelerating the screening of new materials,providing information support and directional guidance for subsequent experimental work.