Instytut Wysokich Ciśnień

Akira KUSABA, PhD in Engineering
Assistant Professor
Research Institute for Applied Mechanics (RIAM)
Kyushu University
Fukuoka 816-8580, Japan
TEL: +81 92 583 7760, FAX: +81 92 583 7936

e-mail: kusaba@riam.kyushu-u.ac.jp
Theory and Machine Learning for Crystal Growth

https://zoom.us/j/99223124275?pwd=MVpNQ2prMGI0azdvUDI3aHZPRnFGQT09

Outline:
Semiconductors are indispensable to the modern information and communication society. Such crystalline materials are manufactured by utilizing crystal growth phenomena. Crystal growth phenomena have been studied experimentally, theoretically, numerically, and more recently by machine learning. This lecture aims to introduce students to theory and machine learning for crystal growth. These two parts are complementary: qualitative understanding and quantitative prediction of the phenomena. In the part of theory, students will learn how classical and analytical theories can be used to understand crystal growth phenomena. After introducing the concept of rate-limiting processes, the formulas for the growth rates limited by nucleation, step flow and mass transport are derived. Also, the need to consider surface reconstruction is discussed. In the part of machine learning, students will learn how machine learning can improve crystal growth experiments. Emphasis is placed on the use of machine learning from the perspective of material scientists and material process engineers.

Content:

Part 1: Theory for Crystal Growth

• Lecture 1 18.04.2023 10.00 - 11.00
• Basic Concept and Early Stage of Growth (Elementary Processes, Thermodynamics, Supersaturation, Nucleation):
  From the elementary processes of crystal growth, the microscopic picture of crystal growth can be acquired. These are also the fundamental concepts for modeling growth phenomena. Basics of thermodynamics, which is important in the modeling as well as energetics and kinetics, is reviewed. After learning about nucleation theory, the relation between growth rate and supersaturation is derived.

• Lecture 2 25.04.2023 10.00 - 11.00
• Atomic Models (Surface Energy, Surface Reconstruction, Surface Phase Diagram, First-principle Calculations, Statistical Mechanics):
  Models that deal with the motion of atoms on simplified surface structures often cannot explain experimental trends. Then, we need to start from creating a surface phase diagram to map the surface structures to the experimental conditions. The surface phase diagram can be created based on the first-principles calculations and statistical mechanics. Once the surface reconstruction is identified, surface energy and migration potential can be incorporated into the model.

• Lecture 3 09.05.2023 10.00 - 11.00
• Mesoscopic Models (BCF Theory, Interplane Diffusion, Monte Carlo Simulations):
  On the surface with flat terraces, the step-flow growth rate can be discussed by the well-known BCF theory. The relation between growth rate and supersaturation in the presence of screw dislocations is derived. Based on a concept similar to BCF theory, an interplane diffusion model can also be created. Crystal growth on rough surfaces, on the other hand, can be studied with Monte Carlo molecular simulations.

• Lecture 4 16.05.2023 10.00 - 11.00
• Macroscopic Models (Thermodynamic Analysis, Driving Force for Growth, Alloy Composition):
  Thermodynamic analysis is used to derive the relation between the growth rate limited by mass transport and experimental conditions. In the industrial chemical vapor depositions, the conditions for the growth rate limited by mass transport are commonly used because of their controllability. This model can also be used to predict the growth feasibility of a new growth method, the alloy composition, and to analyze the effects of film strain by substrate constraint.

Part 2: Machine Learning for Crystal Growth

• Lecture 5 23.05.2023 10.00 - 11.00
• Basic Concept (Regression, Classification, Dimensionality Reduction, Clustering):
  After introducing the basic concepts of machine learning, examples of the applications of various basic machine learning algorithms to materials science will be presented. The visualization of high-dimensional data and the clustering techniques for unlabeled data will be also covered.

• Lecture 6 30.05.2023 10.00 - 11.00
• Bayesian Optimization:
  Bayesian optimization is one of the most used methods in machine learning applications to materials process engineering. This is because your lab experiment or computational simulation is simply considered a black box, allowing for an efficient exploration of the experimental conditions that will achieve your objectives. After understanding how Bayesian optimization works, we will consider how it can be utilized in our research.

• Lecture 7 13.06.2023 10.00 - 11.00
• Multi-objective Optimization (and Data Assimilation):
  In the more challenging optimization requirements that emerge in real-world engineering problems, multiple issues with trade-offs need to be resolved simultaneously. Through examples, we will learn how to use multi-objective optimization in materials process engineering. In addition, data assimilation, in which experimental data improves the predictive performance of simulations, will also be introduced.

• Lecture 8 20.06.2023 10.00 - 11.00
• Summary, Advanced models and Applications:
  This lecture will summarize the important content of the above seven lectures. In addition, more advanced and recent models that could not be introduced before will be presented here. I will also introduce my own research applying crystal growth theory and machine learning.