Educational Material and Training
This document introduces basic concepts of data science and machine learning in the context of materials science applications. The focus is on hands-on activities where readers use open, online tools in nanoHUB to explore the concepts being introduced. Topics covered include querying data resources and data organization and preparation. Regression exercises, including neural networks, to predict materials properties from a set of descriptors and a classification exercise designed to predict the crystal structure of elemental metals. The examples are provided as fully contained Jupyter notebooks and state of the art, open software. They are designed to introduce to the main steps in data science workflows and readers can modify or extend them to solve additional problems.
This course will provide the conceptual foundation so that a student can use modern statistical concepts and tools to analyze data generated by experiments or numerical simulation. We will also discuss principles of design of experiments so that the data generated by experiments/simulation are statistically relevant and useful. We will conclude with a discussion of analytical tools for machine learning and principle component analysis. At the end of the course, a student will be able to use a broad range of tools embedded in MATLAB and Excel to analyze and interpret their data.
Reinforcement learning is a paradigm that aims to model the trial-and-error learning process that is needed in many problem situations where explicit instructive signals are not available. It has roots in operations research, behavioural psychology and AI. Recently Deep Reinforcement Learning methods have achieved significant successes by marrying the representation learning power of deep networks and the control learning abilities of RL. This has resulted in some of the most significant recent breakthroughs in AI such as the Atari game player and the Alpha Go engine from Deepmind. This success has opened up new lines of research and revived old ones in the RL community. In this talk, Professor Ravindran will introduce the reinforcement learning paradigm and motivate the need for deep RL.
Material structures that occur in nature are commonly made up of complex architectures arranged in a hierarchy. These hierarchical architectures are made up of different structural levels consisting of a unique arrangement of simple constituents, acting as building blocks, that satisfy a local demand for certain mechanical properties. One structural level functions as a constituent of the next level in the hierarchy and so on, resulting in materials that can exhibit unique mechanical properties (i.e. different from those of the composing constituents). However, the design of hierarchical materials is challenging due to the enormous size of the design space. For example, in a single structural level, the number of constituent combinations can reach upwards to three orders of magnitude. Furthermore, the number of possible combinations increases exponentially from one structural level to the next. This research puts forward the proposition of utilizing a Neural Network (NN) to identify the optimal arrangements of constituents in a structural level. The hierarchical structure of interest to be studied with the NN is a tape spring ligament made of four different constituents capable of snap-through instabilities and energy dissipation. At this stage of research, we are investigating the most important features of the ligament such that the NN can correctly classify between good and bad structures. Training data is generated using finite element simulations, which are considered as ground truth. After the research group has successfully used the NN to find optimal structures, models will be fabricated with steel via stamping for experimental validation.