Trained Convolutional Neural Network Model Research Articles

Classification of tree forms in aerial LiDAR point clouds using CNN for 3D tree modelling

ABSTRACT Three-dimensional models of trees that correspond to the real-world forms of the trees on the ground are used in urban planning, solar power estimation, and other disciplines. Previous studies have focused on generating 3D tree models from high-density point cloud data such as Terrestrial Laser Scanning (TLS) data, which is expensive and limited to small spatial extents. However, there has been limited exploration of inexpensive solutions to model trees over large spatial extents. The goal of this study is to use widely available discrete return Airborne Laser Scanning (ALS) data along with field-captured tree photographs and Google Street View (GSV) images to develop 3D equivalents of trees over larger spatial extents. To this end, we designed a process to assign representative 3D models for individual trees in discrete return ALS point clouds. This study demonstrates the use of a Convolutional Neural Network (CNN) model and 3D models generated with Structure from Motion (SfM) for the realistic modelling of deciduous non-overlapping trees from discrete return ALS data. We classified and labelled the crown shapes of deciduous trees in a study area into four classes based on GSV images of trees. We delineated and segmented non-overlapping deciduous trees from ALS data and reduced them to 2D images using voxel point counts. Next, we trained a CNN architecture to match the 2D images to the corresponding classes observed from GSV images. For each class, we created a representative 3D tree model using field-captured circumnavigational photos of trees and SfM. To demonstrate 3D visualization using the 3D tree models, we created a 3D visualization of the trees surrounding a parking lot. The trained CNN model from this study can be used to classify non-overlapping deciduous trees from discrete return ALS data and subsequently visualize near-realistic 3D tree models of trees.

Convolutional-neural-network-based DES-level aerodynamic flow field generation from URANS data

The present paper culminates several investigations into the use of convolutional neural networks (CNNs) as a post-processing step to improve the accuracy of unsteady Reynolds-averaged Navier–Stokes (URANS) simulations for subsonic flows over airfoils at low angles of attack. Time-averaged detached eddy simulation (DES)-generated flow fields serve as the target data for creating and training CNN models. CNN post-processing generates flow-field data comparable to DES resolution, but after using only URANS-level resources and properly training CNN models. This document outlines the underlying theory and progress toward the goal of improving URANS simulations by looking at flow predictions for a class of simple, two-dimensional, streamlined profiles (i.e., an NACA0006 airfoil simulated at Mach 0.3 over an angle-of-attack range of −8° to 8°). After several design iterations, two trained CNN models predict the pressure and density fields immediately surrounding the airfoils, and those results compare well to DES ground truth data. Generally, the trained models match the DES resolution for the model training data. However, to accurately predict the flow conditions outside of the training set, other approaches are necessary. For this portion of the research, all computational fluid dynamic calculations use NASA’s fully-unstructured-Navier–Stokes-3D solver. The rest of the software comes from Python libraries within the public domain to encourage follow-on research. These libraries include TensorFlow for machine learning models and the sequential model-based optimization to generate the algorithm configuration for hyperparameter optimization.

Mobile app-based tomato disease identification with fine-tuned convolutional neural networks

Rapid detection and diagnosis of plant diseases is crucial for healthy crop growth and maintaining harvest quality. Many diseases, such as bacterial spot, late blight, and leaf mould, can affect tomato cultivation. In this paper, a mobile app-based system for intelligent disease detection on tomato leaves is developed using convolutional neural networks (CNNs). The study uses transfer learning approaches to fine-tune well-known CNN architectures like AlexNet, ResNet-50, SqueezeNet-1.1, VGG19, and DenseNet-121 on a dataset made up of images of tomato leaf from the PlantVillage dataset as well as self-captured images. Over 95% accuracy is attained by all trained CNN models, with DenseNet-121 leading the pack with a stellar 99.85%. Consequently, DenseNet-121 is proposed for use with a user-friendly mobile application to detect disease in tomato leaves. A robust user interface in English and two important Indian languages makes this application more accessible to a wider user base.

Multi-Dataset Comparison of Vision Transformers and Convolutional Neural Networks for Detecting Glaucomatous Optic Neuropathy from Fundus Photographs.

Glaucomatous optic neuropathy (GON) can be diagnosed and monitored using fundus photography, a widely available and low-cost approach already adopted for automated screening of ophthalmic diseases such as diabetic retinopathy. Despite this, the lack of validated early screening approaches remains a major obstacle in the prevention of glaucoma-related blindness. Deep learning models have gained significant interest as potential solutions, as these models offer objective and high-throughput methods for processing image-based medical data. While convolutional neural networks (CNN) have been widely utilized for these purposes, more recent advances in the application of Transformer architectures have led to new models, including Vision Transformer (ViT,) that have shown promise in many domains of image analysis. However, previous comparisons of these two architectures have not sufficiently compared models side-by-side with more than a single dataset, making it unclear which model is more generalizable or performs better in different clinical contexts. Our purpose is to investigate comparable ViT and CNN models tasked with GON detection from fundus photos and highlight their respective strengths and weaknesses. We train CNN and ViT models on six unrelated, publicly available databases and compare their performance using well-established statistics including AUC, sensitivity, and specificity. Our results indicate that ViT models often show superior performance when compared with a similarly trained CNN model, particularly when non-glaucomatous images are over-represented in a given dataset. We discuss the clinical implications of these findings and suggest that ViT can further the development of accurate and scalable GON detection for this leading cause of irreversible blindness worldwide.

A Pyramid Deep Feature Extraction Model for the Automatic Classification of Upper Extremity Fractures.

The musculoskeletal system plays a crucial role in our daily lives, and the accurate diagnosis of musculoskeletal issues is essential for providing effective healthcare. However, the classification of musculoskeletal system radiographs is a complex task, requiring both accuracy and efficiency. This study addresses this challenge by introducing and evaluating a pyramid deep feature extraction model for the automatic classification of musculoskeletal system radiographs. The primary goal of this research is to develop a reliable and efficient solution to classify different upper extremity regions in musculoskeletal radiographs. To achieve this goal, we conducted an end-to-end training process using a pre-trained EfficientNet B0 convolutional neural network (CNN) model. This model was trained on a dataset of radiographic images that were divided into patches of various sizes, including 224 × 224, 112 × 112, 56 × 56, and 28 × 28. From the trained CNN model, we extracted a total of 85,000 features. These features were subsequently subjected to selection using the neighborhood component analysis (NCA) feature selection algorithm and then classified using a support vector machine (SVM). The results of our experiments are highly promising. The proposed model successfully classified various upper extremity regions with high accuracy rates: 92.04% for the elbow region, 91.19% for the finger region, 92.11% for the forearm region, 91.34% for the hand region, 91.35% for the humerus region, 89.49% for the shoulder region, and 92.63% for the wrist region. These results demonstrate the effectiveness of our deep feature extraction model as a potential auxiliary tool in the automatic analysis of musculoskeletal system radiographs. By automating the classification of musculoskeletal radiographs, our model has the potential to significantly accelerate clinical diagnostic processes and provide more precise results. This advancement in medical imaging technology can ultimately lead to better healthcare services for patients. However, future studies are crucial to further refine and test the model for practical clinical applications, ensuring that it integrates seamlessly into medical diagnosis and treatment processes, thus improving the overall quality of healthcare services.

Deep learning prediction of waterflooding-based alteration of reservoir hydraulic flow unit

The hydraulic flow unit (HFU) is a comprehensive representation of reservoir quality, and the type of reservoir HFU alters dramatically due to the impact of waterflooding. The accurate prediction of HFU characteristic alterations based-waterflooding is crucial but difficult because the alterations are influenced by many factors like geological conditions and production history. In this study, we propose a deep learning-based method using both static geological data and dynamic production history data to predict HFU characteristic alterations waterflooding-based. We classify six HFU types using the k-means clustering of geological variables, including the flow zone indicator (FZI), permeability, and shale volume. The deep-learning model uses the initial static geological variables, including FZI, permeability, and shale volume, and dynamic production data as inputs, whereas the HFU type after waterflooding is used as the target. Convolutional neural networks (CNNs) are adopted for the prediction of HFU characteristic alterations and also compared with support vector machines (SVM) and back propagation neural networks (BPNNs). The prediction accuracy of the CNN model is 93.3% and 87.1% for the validation and testing datasets, respectively, which performs much better than other machine learning methods. In addition, we explore the factors affecting the performance of different models. CNN model can extract useful information from complicated production history data more effectively than other machine learning methods. Finally, the trained CNN model is used to predict the spatial distribution of HFU at different development stages in the study area. In summary, the proposed method can accurately predict dynamic alterations in the reservoir HFU type after waterflooding, which is very useful for the adjustment of oilfield production measures.

Privacy-Preserving Electricity Data Classification Scheme Based on CNN Model With Fully Homomorphism

Data classification of users' electricity consump- tion provides an in-depth analysis for users' electricity con- sumption status, which plays a vital role in the management and distribution of electric energy. So, some data classification methods have been proposed to solve the classification problem of electricity consumption data. However, plaintext-based data classification may bring about the privacy leakage of electric- ity consumption data. In this paper, we propose a privacy- preserving classification scheme for electricity consumption da- ta under fog computing-based smart metering system, which is based on convolutional neural network (CNN) model with fully homomorphic method (CKKS). The target of our proposed scheme is to solve the leakage problem of private electricity consumption data during the classification procedure. In our scheme, an improved K-means-based labeling algorithm is constructed to process historical electricity consumption data, which is used as the sample data to train the CNN classification model by cloud server. Also, the fog nodes are only permitted to obtain the related ciphertext parameters of the trained CNN model, and perform the classification of ciphertext-based electricity consumption data generated by fully homomorphic method. Based on the classical testing data, the experimental results show that our proposed classification scheme can provide the high classification accuracy of electricity data while protecting the privacy of electricity data.

Deep Learning and Long-Duration PRPD Analysis to Uncover Weak Partial Discharge Signals for Defect Identification

This study focuses on improving the defect recognition accuracy under weak partial discharges (PDs) in epoxy resin through phase-resolved partial discharge (PRPD) analysis. The method is to refine the data rather than enhance the algorithm. Two measurement conditions are compared until PRPD pattern saturation is achieved: one-minute and one-hour durations. The PD data specifically target three void types located at different positions within the epoxy material. The aim is to evaluate how the presence of weak PDs at the PD extinction voltage (PDEV) influences defect recognition accuracy. This research sheds light on the potential implications of neglecting the significance of weak PD signals in defect detection. A convolutional neural network (CNN) model is trained using PRPD data recorded at the repetitive PD inception voltage (RPDIV) and tested using the new PRPD data from both conditions recorded from a lower PDIV to a PDEV. The trained CNN model achieves a defect recognition accuracy of 100% for a one-hour duration, highlighting the importance of not neglecting weak PD signals. This emphasizes the significance of extended measurement duration and pattern saturation in capturing and analyzing weak PD signals for an improved defect recognition. This study contributes to the advancement of practical applications by understanding the behavior of the epoxy material and enhancing defect detection techniques.

REAL-TIME HAND GESTURE DETECTION AND RECOGNITION USING CONVOLUTIONAL ARTIFICIAL NEURAL NETWORKS

Real-time hand gesture detection is an interesting topic in pattern recognition and computer vision. In this study, we propose the use of a Convolutional Neural Network (CNN) to detect and recognize hands in real-time. Our goal is to develop a system that can accurately identify and interpret user gestures in real-time. The proposed approach involves two main stages, namely hand gesture recognition and gesture recognition. For stage detection, we use the CNN architecture to recognize hands in the video. We train the CNN model using a dataset containing various hand gestures. Once a hand is detected, we extract the relevant hand region and proceed to the gesture recognition stage. The gesture recognition stage involves training and testing CNN models for different hand signal recognition. We use a hand gesture dataset that contains a variety of common hand signals. The experimental results show that the proposed system can detect and recognize hand movements in real-time with satisfactory accuracy. Although there are still some challenges that need to be overcome, this research provides a solid foundation for further development in real-time hand gesture recognition.

Interpretable deep learning for accelerated fading recognition of lithium-ion batteries

Data-driven approaches have gained increasing attention in the field of battery life-related prediction, as building a comprehensive mechanistic degradation model remains a challenge. Deep learning has emerged as a powerful data-driven fitting method for battery-related applications. However, interpretability remains an issue in this field, hindering the practical utilization of deep learning methods. With the development of interpretable techniques, deep learning methods not only can be conducted as black box tools for fitting, but also for exploring the relationship between external battery data and internal electrochemical changes. In this paper, an interpretable deep learning procedure is proposed and exemplified by accelerated fading point (knee-point) recognition based on an open battery dataset. The Gradient-weighted Class Activation Mapping (Grad-CAM) is conducted to explain the link between the input and output of the trained convolutional neural networks (CNN) model. The trained CNN model possesses deep insight into battery degradation, giving the very first warning when accelerated fading occurs. Through interpretability analysis, it is confirmed that the well-trained model can spontaneously focus on features associated with internal battery degradation and identify some additional features beyond existing human experience. The proposed method can be used to discover the relationship between battery data and degradation mechanism by artificial intelligence in the electric vehicles (EVs) field.

Automatic Digitization of JMA Strong-Motion Seismograms Recorded on Smoked Paper: An Attempt Using Deep Learning

Abstract Using deep learning, an attempt was made to automatically digitize waveforms in scanned Japan Meteorological Agency (JMA) analog seismograms of strong-motion seismographs recorded on smoked paper. We manually traced the waveforms in the scanned images to be used as supervised data for a convolutional neural network (CNN) model using these images. In the analog recording system, ground shaking was recorded on paper with a needle attached to the end of an arm. Because the other end of the arm is fixed on a pivot, the needle moves in an arc around the pivot for large shaking compared to the arm length. We carefully considered the effects of the finite arm length in the analog system and trained the CNNs. To validate the learning results, the trained CNN model was applied to images that were not used for the training. The automatic digitization by our method works fairly well, except for seismograms rich in high-frequency components and those with long-period large amplitudes. Even these images can be well digitized by resizing the image height by, for example, 1/2. To further improve the accuracy of automatic digitization, it would be effective to prepare many additional input and supervised data, and retrain the model. Other training is also necessary for other types of seismographs that have features and characteristics different from strong-motion seismograms.

Automated Categorization of Multiclass Welding Defects Using the X-ray Image Augmentation and Convolutional Neural Network.

The detection of weld defects by using X-rays is an important task in the industry. It requires trained specialists with the expertise to conduct a timely inspection, which is costly and cumbersome. Moreover, the process can be erroneous due to fatigue and lack of concentration. In this context, this study proposes an automated approach to identify multi-class welding defects by processing the X-ray images. It is realized by an intelligent hybridization of the data augmentation techniques and convolutional neural network (CNN). The proposed data augmentation mainly performs random rotation, shearing, zooming, brightness adjustment, and horizontal flips on the intended images. This augmentation is beneficial for the realization of a generalized trained CNN model, which can process the multi-class dataset for the identification of welding defects. The effectiveness of the proposed method is confirmed by testing its performance in processing an industrial dataset. The intended dataset contains 4479 X-ray images and belongs to six groups: cavity, cracks, inclusion slag, lack of fusion, shape defects, and normal defects. The devised technique achieved an average accuracy of 92%. This indicates that the approach is promising and can be used in contemporary solutions for the automated detection and categorization of welding defects.

Investigating transfer learning to improve the deep-learning-based segmentation of organs at risk among different medical centers for nasopharyngeal carcinoma

Purpose: Convolutional neural networks (CNNs) offer a promising approach to automating organ segmentation in radiotherapy. However, variations of segmentation protocols made by different medical centers may induce a well-trained CNN model in one center and may not perform well in other centers. In this study, we proposed a transfer learning method to improve the performance of deep-learning based segmentation models among different medical centers using nasopharyngeal cancer (NPC) data.Methods: The NPC data included 300 cases (S_Train) from one institution (the source center) and 60 cases from another (the target center), divided into a training set of 50 cases (T_Train) and a test set of 10 target cases (T_Test). A ResNet CNN architecture was developed with 103 layers. We first trained Model_S and Model_T from scratch with the datasets S_Train and T_train, respectively. Transfer learning was then used to train Model_ST by fine-tuning the last 10 layers of Model_S with images from T_Train. We also investigated the effect of the numbers of re-trained layers on the performance. The performance of each model was evaluated using the dice similarity coefficient, and it was used as the evaluation metrics. We compared the dice similarity coefficient value using the three different models (Model_S, Model_T, and Model_ST).Results: When Model_S, Model_T, and Model_ST were applied to the T_Test dataset, the transfer learning (Model_ST) had the best performance. Compared with Model_S, the p-values of all organs at risk were less than 0.05. Compared with Model_T, the p-values of most organs at risk were less than 0.05, but there was no significant statistical difference in Model_ST for the brain stem (p = 0.071), mandible (p = 0.177), left temporal lobes (p = 0.084), and right temporal lobes (p = 0.068). Although there was no statistical difference for these organs, the mean accuracy of Model_ST was higher than that of Model_T. The proposed transfer learning can reduce the training time by up to 33%.Conclusion: Transfer learning can improve organ segmentation for NPC by adapting a previously trained CNN model to a new image domain, reducing the training time and saving physicians from labeling a large number of contours.

Automatic Fruit Quality Inspection System Using Image Processing

The demand for high-quality fruits has increased significantly, driven by consumers' growing emphasis on health and nutrition. To ensure consistent quality control and efficient fruit grading processes, an automatic fruit quality inspection system using image processing techniques has been developed. This system leverages Convolutional Neural Network (CNN) algorithms to achieve an impressive accuracy rate of 99% in fruit quality assessment. The proposed system involves a multi-step process starting with the acquisition of high-resolution fruit images. These images undergo pre-processing to enhance clarity and eliminate noise or artifacts. Subsequently, the pre-processed images are fed into a trained CNN model for feature extraction and classification based on learned patterns. The CNN model has been trained on a large dataset of labeled fruit images, enabling it to recognize quality attributes such as colour, size, shape, and defects. The system's evaluation involved a diverse range of fruits, encompassing various species, varieties, and maturity stages. The output demonstrate the system's exceptional accuracy, with a 99% success rate in correctly identifying and grading fruit quality attributes. This automatic fruit quality inspection system offers several advantages, including real-time processing, high efficiency, and reduced labor costs, making it suitable for integration into fruit processing facilities and supply chains. [1]

Numerical modeling of FS-trench IGBTs by TCAD and its parameter extraction method

The electrical behavior of the IGBTs is the key factors to the application of IGBTs. Technology Computer Aided Design (TCAD) is one of the most effective methods for characterizing of electrical behavior of IGBT. However, the great difficulties in extracting the model's parameters hinder its widely applications. In the present work, a high-precision TCAD model for FS-Trench IGBT was established and verified by experiments. Firstly, the parameters necessary for TCAD models were analyzed in theoretical and experimental. The key process parameters need to be determined in a value range table. A Convolutional Neural Network (CNN) was proposed and trained to form a relationship from characteristic curves to key process parameters. Subsequently, the numerical model's parameters were established by combining the trained CNN model with experimental data. Finally, the TCAD model was established and the simulations of IGBTs were carried out and verified by experiments. Results demonstrate that the proposed TCAD model is well consistent with experiments with a Root Mean Square Error (RMSE) of about 3.7 %. This work may provide a feasible method to obtain highly precise numerical models of power semiconductor devices. It may also promote the application of Artificial Intelligence in the device modeling and optimization design.

A Pilot Study: Detrusor Overactivity Diagnosis Method Based on Deep Learning

To develop two intelligent diagnosis models of detrusor overactivity (DO) based on deep learning to assist doctors no longer heavily rely on visual observation of urodynamic study (UDS) curves. UDS curves of 92 patients were collected during 2019. We constructed two DO event recognition models based on convolutional neural network (CNN) with 44 samples, and tested the model performance with the remaining 48 samples by comparing other four classical machine learning models. During the testing phase, we developed a threshold screening strategy to quickly filter out suspected DO event segments in each patient's UDS curve. If two or more DO event fragments are determined to be DO by the diagnostic model, the patient is diagnosed as having DO. We extracted 146 DO event samples and 1863 non-DO event samples from the UDS curves of 44 patients to train CNN models. Through 10-fold cross-validation, the training accuracy and validation accuracy of our models achieved the highest accuracy. In the model testing phase, we used a threshold screening strategy to quickly screen out the suspected DO event samples in the UDS curve of another 48 patients, and then input them into the trained models. Finally, the diagnostic accuracy of patients without DO and patients with DO was 78.12% and 100%, respectively. Under the available data, the accuracy of the DO diagnostic model based on CNN is satisfactory. With the increase of the amount of data, the deep learning model is likely to have better performance. This experiment was certified by the Chinese Clinical Trial Registry (ChiCTR2200063467).

Deep learning-based CNC milling tool wear stage estimation with multi-signal analysis

In this work, the convolutional neural network (CNN), which is a deep learning method in which the features are extracted by an inner process, was performed to detect the wear stages of the milling tool. These stages that define the total lifespan of the tool are known as initial wear (IW), steady-state wear (SSW), and accelerated wear (AW). Short Time Fourier Transform (STFT) was applied to signals, and signal spectrograms were used to train CNN models with different complex architectures. Vibration signals, acoustic emission signals, and motor current signals from The Nasa Ames Milling Dataset were used to obtain the spectrograms. Pre-trained CNNs (GoogleNet, AlexNet, ResNet-50, and EfficientNet-B0) detected the tool wear stage with varying accuracies. It has been seen that the time duration of model training increases as the size of the dataset grows and the network architecture becomes more complex. The recommended method has also been tested on the 2010 PHM Data Challenge Dataset. CNN shows promise for condition monitoring of milling operations and detecting tool wear stage.

A Novel Fake-News Dataset and Detection System to Mitigate Cyber War with Emphasis on Nigerian News Events

Fake-news refers to a cyber-weapon launched through the social media, as, its consequence can result to the breakdown of law and order in the society both physically and on the cyber-social-space. In Nigeria, there is currently no established law that guides the use of social media. Therefore, the rate at which fake-news propagates is alarming. This paper presents a new dataset, with focus on Nigeria’s trending news such as EndSARS and Herdsmen attacks, which was further used to simulate Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) machine learning models to detect fake-news. The data were extracted from twitter using twitter Application Package Interface (API) and from facebook using a scraping tool. The dataset was encoded using Unicode escape function in python to make all characters accessible by the algorithm and tokenised using Global Vectors for Word Representation. The dataset was used to train CNN and RNN models built in python on google colab platform to detect fake-news using accuracy, sensitivity, recall and F1 score as evaluation metrics. Results showed that RNN performed better in terms of accuracy and precision, at 82.34% and 93.19% compared to 81.96% and 79.65% for CNN, F1 scores are approximately the same for both models and CNN performed better than RNN in terms of recall at 98.03% to 50.61% for RNN.

High-accuracy detection of supraspinatus fatty infiltration in shoulder MRI using convolutional neural network algorithms.

The supraspinatus muscle fatty infiltration (SMFI) is a crucial MRI shoulder finding to determine the patient's prognosis. Clinicians have used the Goutallier classification to diagnose it. Deep learning algorithms have been demonstrated to have higher accuracy than traditional methods. To train convolutional neural network models to categorize the SMFI as a binary diagnosis based on Goutallier's classification using shoulder MRIs. A retrospective study was performed. MRI and medical records from patients with SMFI diagnosis from January 1st, 2019, to September 20th, 2020, were selected. 900 T2-weighted, Y-view shoulder MRIs were evaluated. The supraspinatus fossa was automatically cropped using segmentation masks. A balancing technique was implemented. Five binary classification classes were developed into two as follows, A: 0, 1 v/s 3, 4; B: 0, 1 v/s 2, 3, 4; C: 0, 1 v/s 2; D: 0, 1, 2, v/s 3, 4; E: 2 v/s 3, 4. The VGG-19, ResNet-50, and Inception-v3 architectures were trained as backbone classifiers. An average of three 10-fold cross-validation processes were developed to evaluate model performance. AU-ROC, sensitivity, and specificity with 95% confidence intervals were used. Overall, 606 shoulders MRIs were analyzed. The Goutallier distribution was presented as follows: 0 = 403; 1 = 114; 2 = 51; 3 = 24; 4 = 14. Case A, VGG-19 model demonstrated an AU-ROC of 0.991 ± 0.003 (accuracy, 0.973 ± 0.006; sensitivity, 0.947 ± 0.039; specificity, 0.975 ± 0.006). B, VGG-19, 0.961 ± 0.013 (0.925 ± 0.010; 0.847 ± 0.041; 0.939 ± 0.011). C, VGG-19, 0.935 ± 0.022 (0.900 ± 0.015; 0.750 ± 0.078; 0.914 ± 0.014). D, VGG-19, 0.977 ± 0.007 (0.942 ± 0.012; 0.925 ± 0.056; 0.942 ± 0.013). E, VGG-19, 0.861 ± 0.050 (0.779 ± 0.054; 0.706 ± 0.088; 0.831 ± 0.061). Convolutional neural network models demonstrated high accuracy in MRIs SMFI diagnosis.

Full-field prediction of stress and fracture patterns in composites using deep learning and self-attention

An efficient surrogate modelling framework is proposed for full-field predictions of stresses and cracks in composite material microstructures. The framework comprises two sequential convolutional neural networks (CNNs), predicting the elastic stress fields and the local crack maps, respectively. Training and test data are created from high-resolution fracture simulations of randomly generated representative volume elements (RVEs), including geometric variabilities such as fibre volume fraction and porosity. This work shows that the inclusion of a self-attention layer within the network enables the model to capture relevant local and global features, which are important in determining the heterogeneous stress distribution and crack patterns. The performance of the trained CNN models is evaluated with unseen data. The CNN models speed up the full-field predictions by 3 ∼ 4 orders of magnitude compared to the physics-based model. The surrogate model’s accuracy and efficiency are key enables for applications such as multiscale simulation, microstructure optimisation and uncertainty quantification.