Convolutional Neural Networks are fascinating, especially in their materials engineering applications. I spent the 6 months (from Januray to June 2021) studying this deep learning method to assess the possibility of characterizing structures in nature. Nature has provided us with an abundance of resilient designs and our methods of biomimicry have only gotten smarter. Following this idea, I wanted to bridge our current gap in the design of tougher ceramics by extracting the engineering parameters from natural structures. I achieved this by training a convolutional neural network to learn the patterns of bioinspired designs and label the level of disorder in the structure. A honeycomb design, which is perfectly ordered, can be characterized by fundemental design parameters and so can disordered structures. The model I developed has been able to characterize the level of order in a range of natural structures including honeycombs, nacre, and armadillo shells with a notable level of accuracy. I took a picture of the tiling when I visited the Montreal botanical gardens and my model was able to characterize the level of disorder in that. Pretty neat :)
I used OpenCV to pre-process and segment images for training, and Tensorflow’s implementation of Keras in Python to build and train my model. I attribute a lot of what I have learnt and the code base to Professor Markus J. Buehler, during my fellowship in his Predictive Multiscale Materials Design course at MIT as well as to François Chollet in his book on “Deep Learning with Python”. Finally, without the designs developed by Derek Aranguren van Egmond and the supervision of Hamidreza Yazdani Sarvestani and Behnam Ashrafi, I would not have made it this far.
Lets see what else we can engineer from natural structures.
Comments