Prediction and optimization of mechanical properties of composites using convolutional neural networks

Abueidda, Diab W.; Al-Masri, M.S.; Ammourah, Rami; Ravaioli, Umberto; Jasiuk, Iwona; Sobh, Nahil

doi:10.1016/j.compstruct.2019.111264

Cited by 156 publications

(60 citation statements)

References 39 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In this regard, machine learning approaches can establish the structureproperty relationships and thereby determine the properties in a quick and accurate manner. Several researchers have successfully applied machine learning techniques to predict the properties of composite materials and to optimize composite properties [14,15,16]. For the prediction of the modulus, strength, and toughness of composite materials, Abueidda et al [15] trained a 6-layer Convolutional Neural network on 4.3 million computer-generated composite configurations, achieving excellent results with more than 99% predictive accuracy in each property.…”

Section: Introductionmentioning

confidence: 99%

“…Several researchers have successfully applied machine learning techniques to predict the properties of composite materials and to optimize composite properties [14,15,16]. For the prediction of the modulus, strength, and toughness of composite materials, Abueidda et al [15] trained a 6-layer Convolutional Neural network on 4.3 million computer-generated composite configurations, achieving excellent results with more than 99% predictive accuracy in each property. In [16], a 5-layer CNN was proposed and achieved very good predictive results of the effective elastic properties of composites with arbitrary shapes and distribution of inclusions.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

From microstructural images to properties - an interpretable deep learning approach to predict elastic-plastic properties of fiber composites

Ponnusami¹

2021

Preprint

View full text Add to dashboard Cite

The application of machine learning in the field of materials engineering can facilitate materials design and enable faster discovery of novel materials. This paper presents a deep learning approach for the prediction of the transverse elastic and plastic properties of unidirectional fibre reinforced composites directly from images of their microstructures. The training dataset consists of finite element predictions of the elastic-plastic properties of a set of 2D representative volume elements of unidirectional composites with different volume fraction and fiber diameter. Single-regression, Fully-connected Neural Networks (FcNNs) and a multi-regression Convolutional Neural Network (CNN) are designed and trained to predict 5 mechanical properties, using as input the images of the material’s microstructure. The performance of the FcNNs and of the CNN are compared; the CNN is shown to perform the most effective regression, predicting the properties of the composites with an accuracy of 99%. A key value addition in this research lies in the explainability of the otherwise ‘blackbox’ CNN model, guiding the reader through the model’s predictive process. We test the CNN on real, relatively low-resolution micrographs of composite microstructures and we find that this provides accurate predictions.

show abstract

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

From microstructural images to properties - an interpretable deep learning approach to predict elastic-plastic properties of fiber composites

Ponnusami¹

2021

Preprint

View full text Add to dashboard Cite

show abstract

“…However, what if the problem cannot be described by linear regression model or the design space has several distribution patterns to achieve optimality? Diab et al [ 22 ] extended the above work by implementing the GA to conduct the optimization. However, during the optimization, the crossover and mutation process changed the material ratio, so a fixed volume fraction of soft (hard) blocks could not be maintained, which led to their system not being mass-conserved and results being unconvincing.…”

Section: Introductionmentioning

confidence: 99%

A Novel Long Short-Term Memory Based Optimal Strategy for Bio-Inspired Material Design

2021

View full text Add to dashboard Cite

Biological materials have attracted a lot of attention due to their simultaneous superior stiffness and toughness, which are conventionally attributed to their staggered structure (also known as brick and mortar) at the most elementary nanoscale level and self-similar hierarchy at the overall level. Numerous theoretical, numerical, and experimental studies have been conducted to determine the mechanism behind the load-bearing capacity of the staggered structure, while few studies focus on whether the staggered structure is globally optimal in the entire design space at the nanoscale level. Here, from the view of structural optimization, we develop a novel long short-term memory (LSTM) based iterative strategy for optimal design to demonstrate the simultaneous best stiffness and toughness of the staggered structure. Our strategy is capable of both rapid discovery and high accuracy based on less than 10% of the entire design space. Besides, our strategy could obtain and maintain all of the best sample configurations during iterations, which can hardly be done by the convolutional neural network (CNN)-based optimal strategy. Moreover, we discuss the possible future material design based on the failure point of the staggered structure. The LSTM-based optimal design strategy is general and universal, and it may be employed in many other mechanical and material design fields with the premise of conservation of mass and multiple optimal sample configurations.

show abstract

“…However, many technologically important composites are synthesized with a variety of forms and a relatively large volume fraction of reinforcement, which limits the applicability of homogenization methods. In contrast, deep learning techniques have been employed to predict the mechanical properties of two-dimensional checkerboard composites [15,[27][28][29] or three-dimensional linear elastic composites [30,31]. It is well known that with more data, the generalization of a deep neural network (DNN) model is improved for unseen data.…”

Section: Introductionmentioning

confidence: 99%

Transfer learning for enhancing the homogenization-theory-based prediction of elasto-plastic response of particle/fiber-reinforced composites

Jung¹,

Kim²,

Park³

et al. 2021

Preprint

View full text Add to dashboard Cite

We recently proposed the adaptive incrementally affine method as a novel homogenization scheme that enables accurate prediction of the nonlinear mechanical response of particulate-reinforced composites under cyclic or multiaxial loading conditions. However, being a mean-field homogenization method that relies on the solution of Eshelby’s inclusion problem, the aforementioned method does not provide accurate predictions when the shape of the reinforcement has a high aspect ratio or the interaction among reinforcements becomes significant. Herein, we propose a combined theoretical and data-driven approach in which the homogenization method is followed by transfer learning to enhance the prediction of the nonlinear mechanical response of particle/fiber-reinforced composites. We trained a deep neural network (DNN) with a massive stress-strain curve dataset of a composite subjected to uniaxial loading in the elasto-plastic regime based on the adaptive incrementally affine homogenization method; then, we fine-tuned the pre-trained DNN via transfer learning with a relatively small dataset based on time-consuming three-dimensional (3D) finite element analyses (FEA). The transfer learning approach exhibited better predictive performance than the DNN directly trained only with the FEA dataset did, for a wide range of reinforcement volume fractions and shapes. With transfer learning, the DNN exploits the knowledge learned from the homogenization theory to perform a new learning task on a small dataset from 3D FEA. The combined theoretical and data-driven approach proposed herein can be extended to other composite analyses when experiments or simulations require excessive time and cost for a large dataset and the applicability of the homogenization theory is limited.

show abstract

Prediction and optimization of mechanical properties of composites using convolutional neural networks

Cited by 156 publications

References 39 publications

From microstructural images to properties - an interpretable deep learning approach to predict elastic-plastic properties of fiber composites

From microstructural images to properties - an interpretable deep learning approach to predict elastic-plastic properties of fiber composites

A Novel Long Short-Term Memory Based Optimal Strategy for Bio-Inspired Material Design

Transfer learning for enhancing the homogenization-theory-based prediction of elasto-plastic response of particle/fiber-reinforced composites

Contact Info

Product

Resources

About