In this article, we will be discussing the role of CNN to predict the composite properties beyond the elastic limit. how AI is becoming state of the art in all other fields? This article is mainly focuses on mechanical engineering students to understand how AI is use in their subjects.
- Introduction
- Problem Statement
- Material and methods
- Result and discussion
This article is inspire by the research paper “Using convolutional neural networks to predict CNN composite properties beyond the elastic limit” by Charles Yang, Youngsoo Kim, Seunghwa Ryu, and Grace X. Gu (doi:10.1557/mrc.2019.49).
Introduction to CNN Composite Properties:
AI is becoming more and more critical in today’s world and is now extensively use in physics, biology, mechanical engineering, and all other fields more than ever.
In computation tasks, AI is providing better alternatives than simulation software because simulation software takes a lot of time and computational resources for simulating real-world scenarios. Take for example you want to make composite material for lightweight applications like the structure of the drone.
you will be using composite having different layers where the input parameter like layer orientation, thickness, composition plays an important role. for optimizing the output parameter like toughness, elastic properties, strength, etc. Now in simulation software, putting all possible input parameters is a huge set of single tasks for simulation software and your simulation can take a year or more depending on mesh and element size.
Moreover, for a specific set of output, it is not possible to predict input. Here AI helps us in solving these problems. We usually take data from FEM software (or any other simulation software) and our AI model on these data. And our AI model can predict the given task giving it has performed well on the test set
Composite materials are use in a variety of products due to the cheap cost of constituent materials, high strength, lightweight property. And these days, due to the limitations of simulation software, we need AI to solve our real-world problems.
Problem Statement:
In fig. 1, we took 2-D 121 blocks of composite materials containing 70 hard blocks and 51 soft blocks whose material properties. Each block is 1m x 1m and is divide up into 144 elements, totaling 17,424 elements in 121 blocks. The arrangement of blocks is complete randomly and the total possible arrangement is large 1030.
but we take 26000 possible combinations of the possible random blocks for data generation using FEM. A 2.5 m crack is placed in the middle of the entire square 2-D plate and tensile testing completes on the FEM simulation software i.e. ABAQUS. We want to predict the toughness, strength, and elastic modulus of any possible arrangement of blocks with CNN and compare it with other ML algorithms. Note that CNN is computationally faster and cheaper than simulation software for full crack propagation.
Materials & Methods:
Data preprocessing:
Hard block and soft block represents 0 and 1 respectively and are place in row and column as shown in figure 1(b). This form an image of 0 and 1 and it is our input for CNN. Separately tensile testing of the block is complete in ABAQUS whose result is use as trusted results.
CNN filter visualization:
This is a non-convex problem and hence we can use gradient descent or Nesterov accelerated momentum. because for convex problem SGD and NAG performs very badly on or near the saddle point. Here our image is binary, but we treat our problem as continuous as a simplification. The intermediate value between hard and soft blocks will be linear interpolation between 0 and 1.
Baseline Model:
Our baseline models are Linear regression and random forest with 100 decision trees and CNN performance will be compare to these baseline models. thus, For more details, refer original paper. by Charles Yang, Youngsoo Kim, Seunghwa Ryu, and Grace X. Gu (doi:10.1557/mrc.2019.49)
Results and discussions:
Fig.2 shows the histogram of 26,000 composite configuration. Its elastic modulus, strength, and toughness plot against the count and pearson correlation coefficient calculate.
Results:
From fig 3, we can see CNN matches the expected output more closely than random forest and linear regression for predicting modulus, strength, and toughness.
Fig. 4 shows the model performance with dataset size and it is well known (in general) that Deep neural networks performs better than Machine learning algorithm. hence for 95% confidence interval, CNN performs better than linear regression and random forest (for dataset greater than 5000 to 1000)
In Fig. 5, it is shown that when CNN standard layer is changed, what is the effect on prediction? So standard CNN (Base model) is more close to actual data. It is interesting to observe the importance of batch normalisation and hyperparameter like number of neural layers.
Conclusion on CNN Composite Properties:
however, We can conclude that Artificial neural networks can be used with composite material property prediction. Not only this, it is increasingly becoming popular in all kind of engineering applications.
Written By: Himanshu Kumar Singh
Reviewed By: Rushikesh Lavate
If you are Interested In Machine Learning You Can Check Machine Learning Internship Program
Also Check Other Technical And Non Technical Internship Programs
