This study aims to propose a method for selecting parameters of a linear equation solver considering properties of coefficient matrix, analysis method, and computer environments. We propose a learning model that appropriately selects parameters of a linear equation solver, for example types of iterative solvers and preconditioners, by machine learning based on various examples. The proposed method builds a neural network for each combination of iterative linear solvers and preconditonors, and estimates the required number of iterations using the features of coefficient matrix. Estimating the number of iterations, we can obtain the computation time by a simple benchmark for each computer, which can be used as an indicator to select parameters of the linear solver. In order to evaluate the feasibility of the proposed method, an automatic matrixgeneration system was also established. As a numerical example, the generalization performance was evaluated by the cross-validation method.

This paper describes optimization algorithms as numerical methods to find the centroidal Voronoi tessellation (CVT). The CVT can expect to put particles at evenly spaced apart with capturing the curvature of the boundary and then generate an efficient initial particle distribution for the particle methods to analyze the fluid flow problems with slopes and curved walls. However, a problem to find CVT is classified as the NP-hard. Therefore, there is a great demand for an algorithm to solve the problem efficiently. This paper considers finding CVT as optimization problem about finding minimum energy and then applies optimizers in the field of machine learning such as the Momentum SGD and the Adam. Moreover, this paper discusses on hyperparameters of optimizers and performances of iteration counts to find CVT.

Fat content is an important index of the added value of meat and livestock by-products. In this study, CNN was used to classify the normal liver and fatty liver to identify the morphology of chicken liver. To recognize the appearance of chicken liver, the feature extraction method, and the trained deep learning model vgg16 were used for transfer learning. The validity of vgg16 is verified by comparing it with the baseline model without transfer learning.

With advances in computers, observation, and simulation technology, it becomes an era when a large amount of data is generated, and it is becoming more important to find out the meaning and knowledge contained in the data. In this paper, we formulated the process of finding the governing equation describing the given data as a symbolic regression problem. In the proposed method, ”partial differential function” is introduced into Genetic Programming to generate partial differential equations automatically, and the generated equations and data are compared and evaluated to automatically distill equations with less error. We conducted numerical experiments to estimate the governing equation from fluid simulation data and evaluated the validity of the proposed method. As a result, the original equation was obtained with high probability, and it was found that the proposed method becomes an effective tool to find useful modeling to represent the data.

Mesh-structured data is an important data structure to perform numerical analyses such as the finite element method and the finite volume method. It is known that graph neural networks (GNNs) can deal with mesh-structured data since meshes can be regarded as graphs. In this work, we demonstrate GNNs are useful in learning finite element analysis results. The proposed method efficiently leverages spatial information; that is, the input feature does not change under any rotation and translation. We show that our model generalizes to much larger meshes than these in the training dataset. Moreover, our model can perform inference for meshes, which have up to one million nodes.

In this paper, we propose a method for predicting computational fluid dynamics (CFD) results using Convolution LSTM. Convolutional LSTM is a method that combines Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM). In addition, this method can predict future states with high accuracy by holding spatial information and time series. First, Convolution LSTM was trained using the visualization results of CFD analysis (image information). And it showed its usefulness. Next, we performed learning using physical quantities on this learning machine and obtained some prediction accuracy.