Abstract:Amorphous silicon nitride (a-SiN) is a material which has found wide application due to its excellent mechanical and electrical properties. Despite the significant effort devoted in understanding how the microscopic structure influences the material performance, many aspects still remain elusive. If on the one hand \textit{ab initio} calculations respresent the technique of election to study such a system, they present severe limitations in terms of the size of the system that can be simulated. Such an aspect plays a determinant role, particularly when amorphous structure are to be investigated, as often results depend dramatically on the size of the system. Here, we overcome this limitation by training a machine-learning (ML) interatomic model to \textit{ab initio} data. We show that molecular dynamics simulations using the ML model on much larger systems can reproduce experimental measurements of elastic properties, including elastic isotropy. Our study demonstrates the broader impact of machine-learning potentials for predicting structural and mechanical properties, even for complex amorphous structures.

What problem does this paper attempt to address?

The problem this paper attempts to address is: accurately predicting the structure and mechanical properties of amorphous materials (specifically silicon nitride) through machine-learning potentials. Specifically, the researchers aim to overcome the system size limitations of traditional ab initio calculations when simulating amorphous materials, thereby enabling more accurate predictions of the mechanical properties, such as elasticity, of these materials. ### Background - **Silicon Nitride (Si3N4)** is a ceramic material with excellent mechanical and electrical properties, widely used in high-temperature engineering components, cutting tools, protective coatings for nanocomposites, and other fields. - **Amorphous Silicon Nitride (a-Si3N4)** also exhibits many technical advantages in thin and thick film forms, such as high dielectric constant, high energy barrier, radiation resistance, and oxidation resistance. - **Ab initio calculations** are the preferred technique for studying amorphous materials, but they are computationally expensive and limited by system size, making it difficult to accurately simulate the complex structure and mechanical properties of amorphous materials. ### Research Objectives - **Develop machine-learning potentials**: Train machine-learning models (such as Moment Tensor Potentials, MTP) to accurately describe the interatomic interactions in amorphous silicon nitride. - **Large-scale molecular dynamics simulations**: Use the trained machine-learning potentials to perform large-scale molecular dynamics simulations to predict the structure and mechanical properties, particularly the elastic properties, of amorphous silicon nitride. - **Validate model accuracy**: Compare the predictions of the machine-learning model with experimental data and ab initio results to verify the model's accuracy and reliability. ### Main Contributions - **Overcoming system size limitations**: By using machine-learning potentials, simulations can be performed on larger system sizes, leading to more accurate predictions of the properties of amorphous materials. - **Extending application range**: Demonstrating the broad application prospects of machine-learning potentials in predicting the structure and mechanical properties of complex amorphous structures. ### Methods - **Ab initio molecular dynamics (AIMD)**: Generate initial structures and training datasets. - **Machine-learning potentials (MTP)**: Train and optimize the machine-learning model. - **Classical molecular dynamics (MD)**: Use the trained MTP for large-scale simulations to predict structure and mechanical properties. - **Experimental validation**: Measure the density, radial distribution function (RDF), structure factor (S(Q)), and mechanical properties of amorphous silicon nitride to validate the model's accuracy. ### Results - **Structural characteristics**: The radial distribution function (RDF) and structure factor (S(Q)) predicted by the MTP model are consistent with experimental data and ab initio results. - **Mechanical properties**: The elastic constants predicted by the MTP model match the experimental data and ab initio results, especially the performance variations under different cooling rates. ### Discussion - **Impact of model size**: The study examined the effect of different system sizes on the elastic response of amorphous silicon nitride and found that the machine-learning model performs more accurately on larger system sizes. - **Experimental validation**: Experimental results further validated the reliability and accuracy of the machine-learning model. In summary, this paper successfully addresses the system size limitation in predicting the structure and mechanical properties of amorphous materials through machine-learning potentials, providing new tools and methods for the study of amorphous materials.

Accurate prediction of structural and mechanical properties on amorphous materials enabled through machine-learning potentials: a case study of silicon nitride

Predicting Density-Related Mechanical Properties of Amorphous Silicon Nitride Using Molecular Dynamics Simulations

Predicting the Fracture Propensity of Amorphous Silica Using Molecular Dynamics Simulations and Machine Learning

Quantifying Chemical Structure and Atomic Energies in Amorphous Silicon Networks

Machine Learning Potential Powered Insights into the Mechanical Stability of Amorphous Li-Si Alloys

High-Accuracy Neural Network Interatomic Potential for Silicon Nitride

Study on the Mechanical Properties of Beta Silicon Nitride Based on Neural Network Potential

Mechanical properties of single and polycrystalline solids from machine learning

The road to accuracy: machine-learning-accelerated silicon ab initio simulations

Relationship of structure and mechanical property of silica with enhanced sampling and machine learning

Atomistic structure of amorphous silicon nitride from classical molecular dynamics simulations

Modelling atomic and nanoscale structure in the silicon–oxygen system through active machine learning

Modelling atomic and nanoscale structure in the silicon-oxygen system through active machine learning

Prediction of Carbon Nanostructure Mechanical Properties and Role of Defects Using Machine Learning

An Accurate and Transferable Machine-Learning Interatomic Potential for Silicon

Prediction of bending strength of Si3N4 using machine learning

Efficient prediction of mechanical properties of Boron Nitride Nanosheets using artificial neural networks

Quantum-corrected thickness-dependent thermal conductivity in amorphous silicon predicted by machine-learning molecular dynamics simulations

Origins of structural and electronic transitions in disordered silicon

Graph-neural-network predictions of solid-state NMR parameters from spherical tensor decomposition