Improving the accuracy of the neuroevolution machine learning potential for multi-component systems

In a previous paper [Fan Z et al. 2021 Phys. Rev. B, 104, 104309], we developed the neuroevolution potential (NEP), a framework of training neural network based machine-learning potentials using a natural evolution strategy and performing molecular dynamics (MD) simulations using the trained potentials. The atom-environment descriptor in NEP was constructed based on a set of radial and angular functions. For multi-component systems, all the radial functions between two atoms are multiplied by some fixed factors that depend on the types of the two atoms only. In this paper, we introduce an improved descriptor for multi-component systems, in which different radial functions are multiplied by different factors that are also optimized during the training process, and show that it can significantly improve the regression accuracy without increasing the computational cost in MD simulations.

Technology for Cast Iron Microstructure Analysis in SIAMS Software Using Neural Networks

The paper covers key issues of metal and alloys’ microstructure control using cast iron microstructure examples, and ways of resolving these issues by integration of neural networks into algorithms of SIAMS software. Paper lists key specifics of using the technology and training neural network, aimed at improving algorithm reproducibility, analysis acceleration and simplification. The method for training neural network models as part of the SIAMS software includes functionality for assessing the quality of training. The described method allows you control the model error using the value of the loss function. Developed algorithms in form of ready solutions were integrated into the SIAMS software package, and can be recommended for serial microstructure control in industrial laboratories.

Development of synthetic data generation algorithms for training neural network models for detecting objects in an image

The paper is devoted to the development of synthetic data generation algorithms for training models of object detectors in the image. Modern SOTA architectures based on convolutional neural networks, as well as methods for their training, are considered as target models. The features that a training set based on synthetic data must have for the stable operation of the model on a set of natural data are revealed. The proposed methods and principles for generating such data are described. As an accompanying practical example, the problem of detecting commodity items on the shelves of grocery supermarkets is considered, in the context of which the implemented algorithms were tested.

A Study on the Effect of DropConnect to Control Overfitting in Designing Neural Networks

The neural network is an approach of machine learning by training the connected nodes of a model to predict the results of specific problems. The prediction model is trained by using previously collected training data. In training neural network models, overfitting problems can occur from the excessively dependent training of data and the structural problems of the models. In this paper, we analyze the effect of DropConnect for controlling overfitting in neural networks. It is analyzed according to the DropConnect rates and the number of nodes in designing neural networks. The analysis results of this study help to understand the effect of DropConnect in neural networks. To design an effective neural network model, the DropConnect can be applied with appropriate parameters from the understanding of the effect of the DropConnect in neural network models.