Lecture 1: Applications of machine learning to dynamics in disordered systems

Here I will discuss how we have used a machine learning classification task to identify a quantity we call “softness” to quantify the strength of the cage surrounding each particle. I will discuss how we extract physical understanding from analyzing softness, and how we have used softness as the basis for constructing new models and theories for plasticity and glassy dynamics.

Lecture 2: Machine learning concepts for inverse design in soft matter

In order for artificial neural networks to learn a task, one must solve an inverse design problem. What are all the node weights for the network that will give the desired output? The method by which this problem is solved by computer scientists can be harnessed to solve inverse design problems in soft matter. I will discuss how we have used such approaches to design functional disordered elastic and flow networks. I will also show how we can exploit physics to go beyond what is done in computer science and use local rules rather than global gradient descent approaches to learn in a distributed way.

Lecture 3: Persistent homology for learning about learning

Now that we have designed functional mechanical and flow networks, one can ask how the function—a collective many-body property—emerges from the interaction of modifications that were imposed by the learning process. We have found that persistent homology is a powerful tool for understanding how the ability emerges to perform a basic task inspired by protein allostery. I will describe how we use persistent homology in this context and what it teaches us.