Introduction
Machine and Deep Learning have gained a huge traction in recent decades due to the huge advancements in computing power. It has also become popular since the modern day is data-driven where there is a large amount of data available. The ML/DL models are mainly used to make predictions, classification, forecasting of the responses, and so on.
Currently, I am primarily focused on the application of ML/DL algorithms to engineering problems such as composite structures. I am implementing clustering algorithms such as k-means and spectral clustering, support vector machines for classification, multi-layer feed-forward neural network for regression/classfication, and long-short term memory based recurrent neural network for forecasting of dynamic responses in my research.
Some of the topics correctly being implemented are:
Currently, I am primarily focused on the application of ML/DL algorithms to engineering problems such as composite structures. I am implementing clustering algorithms such as k-means and spectral clustering, support vector machines for classification, multi-layer feed-forward neural network for regression/classfication, and long-short term memory based recurrent neural network for forecasting of dynamic responses in my research.
Some of the topics correctly being implemented are:
- Clustering and classifier based uncertainty quantification of discontinuous responses
- Long-short term memory for forecasting of the responses of composites
- Manifold learning and autoencoders for dimensionality reduction
GSA of discontinuous responses
While the surrogate modeling approaches such as Kriging, PCE, and so on are able to construct an approximate response surface for a continuous function, building an approximate model for discontinuous response function is challenging. To tackle this issue, I have implemented a clustering based approach to find the local regions separated by discontinuity automatically, which are then used to construct a classifier based boundary separation for the discontinuity. This allows for the construction of local surrogate models locally, which can be then used as an ensemble to obtain a global surrogate for UQ of discontinuous responses.
Long-short term memory based fatigue analysis of composites
Estimating the responses of a composites to dynamic loading is computationally expensive while performing finite element simulations. Experiments are also expensive because of the large number of specimens and number of
cycles to failure during the experiment.
The LSTM based approach for forecasting the responses will be beneficial in predicting the future responses of composites and for fatigue life estimation. If incorporated with the experimental data, this will further
reduce the need for lengthy experiments.
cycles to failure during the experiment.
The LSTM based approach for forecasting the responses will be beneficial in predicting the future responses of composites and for fatigue life estimation. If incorporated with the experimental data, this will further
reduce the need for lengthy experiments.
Similar to the implementation of multivariate LSTM for Lorentz system, LSTMs can be applied for forecasting the dynamic responses such as displacements, stresses, and failure behavior of composite structures. It can be implemented in both the computational or experimental analysis in the fatigue life prediction of composite structures as shown below.
Dimensionality reduction using Autoencoders
Furthermore, the neural network based architecture also known as Autoencoders can be applied for dimensionality reduction. AEs are unsupervised learning approach to find the latent space. The latent space in autoencoders is captured in the bottleneck region which is a hidden layer with less number of neurons than the input layer and output layer. By using AEs, one can transform the data with a large number of features to that with a low number of features in the bottleneck region space, z.
Currently, I am exploring the implementation of AE's for dimensionality reduction in the domain of composite structures for efficient UQ while considering multi-responses.