Search

The transverse strength of unidirectional carbon fiber reinforced polymer (UD-CFRP) composites is a high dimensional and nonlinear function of microstructure due to the wide scatter in mechanical properties and complex failure mechanisms, which is a challenging task to develop a general microstructure dependent strength criterion (MDSC) in theory or computation. Volume fraction and distribution of fibers are among the crucial influencing factors. A computational micromechanics and machine learning (ML) combined method is employed to uncover the transverse mechanical response of UD-CFRP composites. High-throughput finite element analyses (FEA) are performed to obtain the transverse behaviors of composites with varying fiber distribution and volume fraction under different loading states. They showed that fiber distribution has different effects on strengths in different failure modes, while the failure modes are closely related to loading states and fiber volume fractions. An ML model is then trained to characterize the relations between composite microstructure and composite strength. Then, the transverse strengths of 1000 new microstructures are predicted, which shows good agreement with FEA results, so that the MDSC of UD-CFRP is constructed by fully accounting for the influence of fiber distribution. Reliability of the method is verified by considering composites with various fiber volume fractions. [Display omitted] • High-throughput finite element analyses are performed to uncover the transverse mechanical response of UD-CFRP composites. • The transverse behaviors of composites with various fiber distribution and fiber volume fraction are presented. • The complex relations between microstructure of composites and transverse strengths are represented. • The microstructure dependent transverse strength criterion is constructed by considering the effect of fiber distribution. [ABSTRACT FROM AUTHOR]

• Unique dataset from Norway's largest dairy company. • Analysis based on domain knowledge and machine learning models. • Identification of important dairy herd features in forecasting milk deliveries. • Machine Learning models performs well in forecasting milk deliveries. • Discussion of models' property for short- and long-term forecasts. Accurate forecasts of milk deliveries to dairy are crucial to secure a stable milk supply to customers, the effective use of raw milk, and the reduction of logistics costs and greenhouse gas emissions. The primary aim of this study was to improve the prediction of the total milk supply from Norwegian dairy farmers to dairies, and to do so by utilizing a traditional time series model, as well as several modern statistical models and machine learning algorithms. The secondary aim was to explore the usefulness of dairy herd features as explanatory variables to forecast the target variable milk delivery to dairy. Several statistical learning models and machine learning algorithms were applied for forecasting, in addition to a traditional time series model and qualitative domain knowledge. Historical data of monthly milk deliveries were collected from Norwegian dairy companies, together with detailed dairy farm herd records. The herd records contain features or explanatory variables, such as the number of cows and inseminations that can influence the response variable – the monthly milk deliveries. Our results show that some of the models applied can predict monthly milk deliveries to dairies with a mean absolute prediction error of 1–2% up to 24-months ahead. While a generalized linear model performed the best up to 3-months ahead, the machine learning models performed the best up to one year ahead. Two years ahead a traditional seasonal autoregressive model performed the best. Because some of the models select the most important features to predict milk delivery, our analytical results also provide valuable information about which features are the most important to predict future milk delivery. [ABSTRACT FROM AUTHOR]

• A method coupling FEA and machine learning is derived. • The predicted data by ANN and RF model are well consistent with the curves of exact values. • Accuracy of predicted result by ANN model is higher than that by RF model. In order to predict mechanical properties of composite laminate, a method coupling finite element analysis (FEA) and machine learning is established to analyze three examples of composite laminates, such as failure factor of Puck theory under random stress state, failure factor and critical buckling eigenvalues of open-hole laminate. By writing Abaqus script, parametric FEA models of composite laminates are built and large samples are generated. Artificial neural network (ANN) model and random forest (RF) model are set up to learn these virtual samples to predict their corresponding mechanical properties for these examples. By analyzing predicted results and FEA values of these mechanical parameters, predicted data are well consistent with the curves of FEA values and the calculated root-mean-square errors are much smaller, which proves this FEA and machine learning coupled method is effective. Even if these errors predicted by ANN model are smaller than RF model, and the learning processes of RF model take less time, these methods with ANN model and RF model are all appropriate to predict mechanical properties of composite laminates. [ABSTRACT FROM AUTHOR]

Determining the allowable compression load (ACL) of notched composite laminates (NCL) requires consideration of buckling, intra- and inter-laminar damage, which is a costly and repetitive task due to the high dimensional and nonlinear relation of ACL to the numerous design variables. In this work, a high-throughput finite element analysis (FEA) and machine learning (ML) combined approach is proposed to predict ACL of laminates. First, the high-throughput FEA model of NCL covering all of the design variables is generated, and the corresponding critical compression loads in terms of buckling, intra- and inter-laminar damage are obtained. Then, ML models are trained, with inputs being the design variables of NCL and outputs being the critical compression loads. ACL and initial failure mode are determined by the minimum compression load of the three failure modes. Moreover, transfer learning is introduced to laminates with new design variables of new ply number, notch shape and laminate type. By fine-tuning the pre-trained ML model using scarce new data, the fine-tuned models are suitable for new laminates. Consequently, ACL and initial failure mode of notched laminates with various design variables are predicted. [ABSTRACT FROM AUTHOR]

Failure assessment is one of the fundamental tasks for optimization of composite pressure vessels (CPVs). However, the extensive design space of composites usually leads to costly and repetitive work of failure assessment that hinders the design and optimization of CPVs. In this work, a combined FEM simulation and machine learning approach is used to predict the failure factors R of CPVs for further optimization. CPVs with various design parameters are automatically generated and analyzed in ABAQUS to obtain failure factors under various external loadings. Then, an assembled deep neural network (DNN) namely Multi-DNN is trained to map relations between design parameters and R of CPVs. Further, transfer learning (TL) is introduced to improve the extensibility of our approach by fine-tuning the pre-trained Multi-DNN using scarce data of new design space. Such a TL-Multi-DNN model can precisely predict R of CPVs with new design parameters. Additionally, combining the trained models and genetic algorithm, the optimization of CPVs is carried out with high efficiency and low computational cost. As a result, R of CPV with various given and new design parameters can be estimated directly by the trained models and the optimal layer sequence of CPV can be obtained efficiently. [ABSTRACT FROM AUTHOR]