Controlling and observing single crystal growth in laboratories has many applications from manufacturing turbine blades to semiconductors. The advancement in the computational power of the modern day computers owes in large to the advanced material engineering techniques.
Growing single crystals is one such innovation, which is labour intensive and time-consuming.
In what might sound as not so obvious in the present day, the researchers of Chinese Academy of Sciences have used machine learning to assess the factors determining the growth of single crystals.
Why Is It Difficult To Grow Crystals
Popular crystallizing techniques such as vapour diffusion, slow cooling, slow evaporation have been put to great use after years of research and most often than not trial and error based experiments.
Theoretically, crystallization should start when the concentration of a compound in a solvent is higher than the solubility product of this compound. Generally, however, crystallization is kinetically hindered and crystals grow only from supersaturated solutions. There are several ways to achieve this metastable state of supersaturation.
The easiest is increasing the concentration by evaporation of the solvent until crystallization sets in. Crystal growing depends on many factors. To begin with, there are parameters like raw material ratio, flux, maximum temperature, minimum temperature, cooling rate, maximum temperature residence time and physicochemical properties such as elemental electronegativity, atomic radius, elements melting point, elemental volatility, the position of the atom in the periodic table among many others.
This problem resonates with something that is prevalent in many machine learning models- the curse of dimensionality. As the number of deciding factors to analyse increase, the complexity involved in achieving a satisfactory outcome increases. And, modern computing machines(along with on-premise data centres, cloud etc) coupled algorithmic advantage is built to do exactly the same; churn complexities and make sense out of data.
How Does ML Help
The researchers used data like growth temperature curves, raw elemental compositions and ratios, and growth conditions.
The authors in this paper, experimented with support vector machine (SVM), decision trees, random forests and gradient boosting decision tree to analyse which of the above factors really make a difference in crystal growth.
Accuracy, f1-scores, recall rates were used for successful sample predictions to evaluate this model.
The results show that:
- The SVM method is relatively stable and works well, with an accuracy of 81% in predicting experimental results. By comparison, the accuracy of laboratory reaches 36%. The decision tree model is also used to reveal which features will take critical roles in growth processes.
- 10-fold cross-validation was used to analyze the model. The single SVM model used to predict experimental results has an accuracy of 80% in describing all the reaction types in its test-set data. The average accuracy over 10 training/validation split is 73%.
Future of ML In Physics
Earlier this year, researchers have demonstrated the use of deep learning for building nuclear reactors where an AI approach was undertaken with a greater focus for the Tokamak reactor, a doughnut-shaped machine that holds hot plasma using a powerful magnetic field.
High dimensional data like the temperature of electrons as a function of radius in the plasma from previous fusion experiments is fed into the Fusion Recurrent Neural Network(FRNN).
The deep learning networks are supplemented with the NVIDIA’s V100 GPUs as the task is computationally intensive and would require high-performance computing clusters. From condensed matter physics to plasma physics, deep learning methods are proving to be more efficient and economical than the conventional methods.
Know more about this work here