​Authors | ​ |
​Nilesh Barla | Published Feb 11, 2021 |
Materials designing and engineering is a very intrinsic job, something that requires a lot of patients and preciseness. Over the years, material designing has become a very crucial part of human civilization not only because it provides structure to our everyday life — from kitchen utensils to buildings, and even from medical to military applications — but also because it provides an understanding of how the universe works through the interaction of the atoms and molecules.
Material science also expands our knowledge by understanding the fragmental design of the material. This understanding helps us to build specific materials based upon our needs and requirements and inspires us to build architectures that create a homogeneous network that bridges humans with nature.
Understanding the material helps us to understand the working of the universe.
We know that the interaction between the atoms and molecules under specific conditions give rise to a certain element or material. What if we could decipher the formation of those materials? When we study and understand the fundamentals and the working of these phenomenons we can build materials that have enhanced properties and structure, those materials can have a spectrum of alloys, high entropy alloys, biomaterials, nanomaterials, shape memory alloys, bulk metallic glass et cetera.
By definition "materials science is a syncretic discipline hybridizing metallurgy, ceramics, solid-state physics, and chemistry."
There are of course different methods and techniques that have been developed to design new materials but these methods usually take a long time. For instance, researching and developing a turbine blade for aerospace purposes could take nearly a decade or more and could cost over 100 million dollars. We have to find an alternative way that is quicker, reliable, and that could design materials in less time and reduce cost.
Deep Learning and its implementation
In recent years the idea of deep learning is penetrating through traditional ideologies and proving to be the solution that everyone looking for. Advancements in computer vision and language modeling have led to the creation of generic algorithms that have defeated humans in the complex game of Go, able to generate text-sentences with human-level accuracy, predicting protein folding to better understand protein sequencing, et cetera.
Deep learning can provide two types of solutions when it comes to materials design and engineering, one being prediction and the other being generation. Both are useful and can provide new insights into this area.
Prediction
Prediction can provide vital information of what going to happen in a material. For instance, Wang et al. developed a deep learning model that could predict stress with fracture propagation in brittle materials. In brittle materials, the cracks are unstable and that can be very dangerous because it propagates without any additional stress or even after the load is lifted. These brittle fractures cause catastrophic events because of the high internal stresses present in these materials, because of these internal stresses it becomes easy for these cracks to propagate. Simulating these materials can be a tedious process but with enough data and the right deep learning architecture, Wang et al. developed an accurate model that could predict stress in a brittle material.
They implemented a hybrid network that uses Temporal Independent Convolutional Neural Network (TI-CNN) and Bidirectional Long Short-Term Memory (LSTM) Network to capture the spatial features of fractures like the fracture propagation and temporal features respectively. The hybrid network provides accurately correlates and predicts the relationship between initial stress data and the sequence of maximum internal stress before fracture.
In another example, Feng S et al. proposed a transfer learning technique to predict phase formation in materials. In the proposed method the deep learning model inherits parameters from a pre-trained model trained in a different dataset. As it turned that the transfer learning model performed quite well when it was trained in a smaller dataset. The model uses Convolutional Neural Network as the primary architecture. It has the ability to extract spatial features in 2-D structures which gives it an edge over other architectures. These spatial features are then passed through a fully-connected layer which eventually predicts the phase formations.
The network proposed by Feng S et al. could predict BCC, HCP, FCC, amorphous, and mixture with accuracy above 94% in fivefold cross-validation.
Generation
Deep learning algorithms can be used to design new and unknown materials with extraordinary properties. Zhou et al. came up with a neural network proposal that could explore the phase design for high entropy alloys. They built an artificial neural network algorithm and were able to derive and extract important features or representations which enabled the quantitative assessment of how to tune a design parameter for the formation of a certain phase, such as solid solution, intermetallic, or amorphous phase. This exploration led to design parameters that were previously unknown for phase designing in high entropy alloys or compositionally complex alloys.
Materials composition can be generated with deep learning architecture provided that the data fed into the algorithm is legit. The algorithm will follow all the laws and rules example Hume-Rothery Rule, Gibb's Phase Rule, Entropy et cetera only if the data is curated properly.
Material Sequencing is another great example of generating new material design. The idea of sequencing was inspired by Natural Language Processing where a language like English is fed into the deep learning architecture called the Seq2Seq architecture. In this architecture, each sentence is broken into the word vector and then each word vector is assigned with a timestamp that tracks the word sequence while training. During the inference, this model can predict the next immediate word and form a sentence. This type of learning is also called in-context learning that GPT-3 is currently using.
The same idea can be extrapolated in material science as well. If the model is provided with proper data that contains all the jargon containing material science like the atomic number, phase rules, crystal structure, melting point in a particular order then Seq2Seq can generate a new composition. But this is just an idea but worth trying.
Conclusion
A lot of experimentation is going in this area. Sooner with the availability of more data more complex deep learning models can be developed which can perform with greater accuarcy. The aim of this article was to present a glimpse of what is happening in the material science world in terms of AI.
References
Wang, Y., Oyen, D., Guo, W.(. et al. StressNet - Deep learning to predict stress with fracture propagation in brittle materials.
Chao Jiang, Gao-Feng Zhao, Nasser Khalili Than - On crack propagation in brittle material using the distinct lattice spring model
Feng, S., Fu, H., Zhou, H. et al. A general and transferable deep learning framework for predicting phase formation in materials.
Zhou, Z., Zhou, Y., He, Q. et al. Machine learning guided appraisal and exploration of phase design for high entropy alloys