Image Credit: WAYHOME studio/Shutterstock.com
Artificial Intelligence (AI) has been touted as the fourth pillar of science by the NVIDIA CEO Jensen Huang. The immense significance it has shown in materials science in recent years only strengthens the idea.
ML is a branch of AI that consists of a set of computer algorithms that systematically help to create models that can efficiently learn from past data/situations. Materials science or solid-state physics is plagued by the ‘curse of dimensionality’. This means even for a reasonable number of constituents of a material, the structure-property relationships are determined by a high-dimensional parameter space. These relationships are not something trivial to be elucidated only by experiments or theory. This is precisely where the applications of AI or machine learning (ML) come in to play.
ML Predictions in Materials Science
Accelerated and accurate predictions of phase diagrams, crystal structures, materials properties, development of interatomic potentials and energy functionals for increasing the speed and accuracy of materials simulations or on‐the‐fly data analysis of high‐throughput experiments are some examples of domains of materials science where there has been a significant application of ML in recent times.
Data Sources for New Materials Design
Experiments have traditionally remained the primary procedure to find, characterize or design new materials. However, experimental accuracy is highly demanding with respect to resources, equipment and time. The introduction of various computational methods in materials science, such as the density functional theory (DFT), Monte Carlo simulations, and molecular dynamics, enabled researchers to explore the phase and composition space more efficiently than ever.
A combination of experiments and computer simulations has contributed towards exploring materials design more economically. High-throughput studies of large material groups have become much more accessible thanks to the tremendous growth in computing power complemented by the development of efficient codes. This is an essential step to screen for the ideal experimental candidates. These large-scale computational calculations, in combination with high-throughput studies in experimentation, have produced a vast amount of data, which makes the usage of ML methods possible in the domain of materials science.
ML Methods in Materials Science
ML has various kinds of algorithms. The two main types which have applications in materials science are ‘Supervised Learning’ and ‘Unsupervised Learning’ methods. If the available data has a set of input and output values, supervised learning is used. If the dataset only contains input values, unsupervised learning is used.
Supervised learning works on the available data to create a new mathematical function/model that can effectively predict a new set of output data when applied to a new collection of input data. On the other hand, unsupervised learning helps predict patterns.
There is then the semi-supervised learning, which lies halfway between supervised learning and unsupervised learning methods. In this case, the ML algorithm is provided with both unlabeled as well as labeled data. These techniques are particularly useful when available data are incomplete and for learning new representations.
There is also the reinforcement learning method where, instead of training the model on how to give precise predictions, reinforcement signals from the environment are used to evaluate the quality of the predictions generated and improve the strategies for adapting to the environment.
In the case of materials science, supervised learning methods are mostly implemented.
Supervised and Unsupervised Learning Examples in Materials Science
- Regularized least-squares
- Support vector machines
- Kernel ridge regression
- Neural networks
- Decision trees
- Genetic programming
Selected material applications
- Prediction of processing structure-property relationships
- Development of model Hamiltonians
- Prediction of crystal structures
- Classify crystal structures
- k‐Means clustering
- Mean shift theory
- Markov random fields
- Hierarchical cluster analysis
- Principal component analysis
Selected materials applications
- Analysis of composition spreads from combinatorial experiments
- Analysis of micrographs
The Different Steps of an ML Workflow
Data processing is an essential step of machine learning that directly affects the performance of the machine learning model used for predictions in materials science. Data processing usually consists of two parts: data selection and feature engineering.
- Data Selection: In this step, data is selected comprehensively with respect to its type, quality, and format.
- Feature Engineering: Feature engineering helps to extract suitable features from raw data obtained from experimentation or simulations to enable the application of ML algorithms in materials science.
Modeling comes after data processing. It includes selecting appropriate algorithms, training, and making accurate predictions. There are several algorithms available to implement the four different types of ML methods discussed earlier. They can be divided into two types:
- Shallow Learning: Shallow learning methods include support vector machine (SVM), decision tree (DT), and artificial neural network (ANN). They are mainly used in linear classification .
- Deep Learning: Deep learning (DL) uses a non-linear cascade processing unit for automatic feature extraction. It derives low‐level features to obtain more abstract high‐level representations of attribute categories. DL models typically outperform SL models for non-linear tasks. Examples of DL models used in materials science: convolutional neural network (CNN), recurrent neural network (RNN), deep belief network (DBN), and deep coding network.
There is always a need to validate the stability of an ML model. Model validation is conducted post-training of the model to evaluate the accuracy of the model by using new unseen data, which differs from the data in the training data set. ML methods usually divide the original data into a training set and a test set, and use the training set for model training and the test set for model validation.
Video Credit: TEDx Talks/YouTube.com
Future Research into AI in Materials Science
Future research on applications of AI in materials science is expected to have two directions. The first will be to continue the development of more sophisticated machine learning methods and their applications in materials science. The verification of the usability of machine learning models will be the goal of the second direction
Find out more about how artificial intelligence has been applied to other areas of science.
References and Further Reading
Yue Liu, et al. (2017) Materials discovery and design using machine learning. Journal of Materiomics. https://doi.org/10.1016/j.jmat.2017.08.002
Abdul Hamid Halabi. (2019) NVIDIA CEO Ties AI-Driven Medical Advances to Data-Driven Leaps in Every Industry. [Online] NVIDIA. Available at: https://blogs.nvidia.com/blog/2019/04/09/world-medical-innovation-forum-2019/ (Accessed on 26 March 2020).
Tim Mueller, et al. (2016) Machine Learning in Materials Science: Recent Progress and Emerging Applications. Reviews in Computational Chemistry. https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=915933
Jing Wei, et al. (2019) Machine learning in materials science. InfoMat. https://doi.org/10.1002/inf2.12028
Jonathan Schmidt, et al. (2019) Recent advances and applications of machine learning in solid-state materials science. npj Computational Materials. https://doi.org/10.1038/s41524-019-0221-0