Source-linked AI summary
Prediction and optimization of mechanical properties of composites using convolutional neural networks
Diab W. Abueidda, Mohammad Almasri, Rami Ammourah, Umberto Ravaioli, Iwona M. Jasiuk, Nahil A. Sobh
TL;DR
The paper addresses the challenge of efficiently predicting and optimizing the mechanical properties of two-dimensional checkerboard composites across material compositions and microstructures. It develops a CNN-based quantitative predictor and integrates it with a genetic algorithm, reporting promising model capabilities and optimized composite configurations.
Problem
Efficiently identifying composite configurations with desired mechanical properties is difficult because large design spaces require many costly numerical simulations.
Method
The paper develops a CNN for quantitative composite-property prediction and integrates it with a genetic algorithm for optimizing configurations.
Results
The model shows promising capabilities for quantitative prediction, and it is tested on a dataset separate from the ground-truth training data.
Takeaways & Limitations
CNN-based prediction can be combined with genetic-algorithm optimization to identify composite configurations with improved performance.
Takeaways & Limitations
The reported linear model does not quantitatively predict checkerboard-composite mechanical properties with high accuracy.
Abstract
from arXiv · showhide
In this paper, we develop a convolutional neural network model to predict the mechanical properties of a two-dimensional checkerboard composite quantitatively. The checkerboard composite possesses two phases, one phase is soft and ductile while the other is stiff and brittle. The ground-truth data used in the training process are obtained from finite element analyses under the assumption of plane stress. Monte Carlo simulations and central limit theorem are used to find the size of the dataset needed. Once the training process is completed, the developed model is validated using data unseen during training. The developed neural network model captures the stiffness, strength, and toughness of checkerboard composites with high accuracy. Also, we integrate the developed model with a genetic algorithm (GA) optimizer to identify the optimal microstructural designs. The genetic algorithm optimizer adopted here has several operators, selection, crossover, mutation, and elitism. The optimizer converges to configurations with highly enhanced properties. For the case of the modulus and starting from randomly-initialized generation, the GA optimizer converges to the global maximum which involves no soft elements. Also, the GA optimizers, when used to maximize strength and toughness, tend towards having soft elements in the region next to the crack tip.
1. Introduction
Composite-material design seeks robust mechanical properties, but exhaustive and simulation-based optimization becomes impractical in large microstructural design spaces. This paper addresses prediction and optimization by developing a CNN that quantitatively predicts properties across volume fractions and integrates with a genetic algorithm.
- Motivation: Composite materials are pursued to meet modern engineering requirements for robust mechanical properties.
- Design challenge: Brute-force and topology-optimization approaches become computationally costly because they require many numerical simulations across large design spaces.Individual simulations may take seconds to hours, while gradient-based methods also face difficulties finding gradients.
- Prior approaches: Machine learning has been applied to predict structural and material behavior and search for materials with optimal designs.
- Present work: The paper develops a CNN that quantitatively predicts composite properties over the entire volume-fraction space and combines it with a genetic algorithm for optimization.The optimization considers both constituent volume fractions and their spatial distribution in the microstructure.
2. Methodology
The study combines finite element analysis, statistical sampling, convolutional neural networks, and genetic algorithms to predict and optimize the mechanical properties of cracked 2D checkerboard composites. The methodology evaluates modulus, strength, and toughness for soft–stiff microstructures across 8x8 and 16x16 grids.
- A CNN quantitatively predicts the modulus, strength, and toughness of 2D checkerboard composites containing soft ductile and stiff brittle materials.
- 2.5. Genetic algorithm: A genetic algorithm uses selection, crossover, mutation, and elitism to search for composite configurations with optimized mechanical properties.
- 2.1. BVP and FE analysis: Finite element analyses generate ground-truth properties for cracked composites under linear elasticity, isotropy, small deformation, symmetry, and plane stress assumptions.The model uses an edge crack and stops simulations when the crack-tip von Mises strain reaches the corresponding material’s failure strain.
- 2.2. Data description and processing: Central-limit-theorem sampling and Monte Carlo analysis determine dataset sizes and assess population-mean convergence for 8x8 and 16x16 grids.The resulting datasets contain 4.3 million points for the 8x8 grid and 4.9 million for the 16x16 grid.
- 2.2. Data description and processing: The two-material microstructure is encoded as a binary grid, with each configuration labeled by its modulus, strength, and toughness.Rows represent distinct configurations, while binary entries identify stiff and soft material placement.
3. Results and discussion
The linear baseline was inadequate for quantitative prediction, while the CNN accurately predicted checkerboard-composite properties and showed little overfitting. Coupled with a genetic algorithm, it identified high-performing microstructures for single and multiple objectives.
- Linear model: For the 16x16 grid, the linear model’s MAPE was 24%, 32%, and 127% for modulus, strength, and toughness, respectively.The corresponding R2 was 0.928, but errors remained too high.
- CNN model: Training and testing MSE losses converged to very low values, with a tiny post-training difference indicating no critical overfitting.The convergence results apply to all properties across the 8x8 and 16x16 grids after 200 epochs.
- Genetic-algorithm optimization: Multi-objective optimization balanced competing properties, reducing modulus while producing 23.4%–32.8% soft-material fractions.Configurations with high strength and toughness tended to place soft material next to the crack tip.
4. Conclusions
The study develops and validates a CNN that quantitatively predicts modulus, strength, and toughness in 2D checkerboard composites, then integrates it with a genetic algorithm to identify high-performing configurations.
- The CNN quantitatively predicts modulus, strength, and toughness of 2D checkerboard composites.
- Finite-element results provide ground-truth training data, while an unseen dataset is used to test model validity.
- The model shows promising capabilities for structural and materials analysis.
- The developed CNN is integrated with a genetic algorithm optimizer to obtain composite configurations with improved properties.
- CNN models may accelerate current optimization techniques and influence structural and materials design.