COMPUTATIONAL SOLID-STATE PHYSICS

Specialization / Direction

Upon successful completion, students will be able to: use the basic methods of computational materials science: Artificial Intelligence in materials science, Machine learning, Materials data & FAIR principles, Crystal representations (composition–structure–property),Generative Artificial Intelligence (LLMs, diffusion models)

The course introduces research tools for materials theory and simulation and is designed for postgraduate students. It covers how composition–structure–property relationships in materials can be encoded for machine learning, then guides participants through building, training, and evaluating models using open-source tools and public datasets. The course opens with an introduction that clarifies its motivations and scope, sketches a brief history of Artificial Intelligence (AI) in science, and sets expectations and assessments, capped by a hands-on exercise in scientific programming. We then establish machine-learning basics—core concepts and terminology and the three paradigms of learning by example (supervised, unsupervised, reinforcement)—applied immediately in an exercise on predicting crystal hardness. Next, we focus on materials data: surveying data sources and types, the FAIR data principles, and data quality control, before a practical on data-driven thermoelectrics. Representations of crystals follow—composition, structure, and crystal graphs—paired with an exercise in navigating crystal space. Classical learning methods (k-nearest neighbours, k-means clustering, decision trees and extensions) are introduced and tested through a “metal or insulator?” classifier. We then progress to deep learning, moving from the neuron to the perceptron, through network architectures and training, to convolutional neural networks, with an exercise on learning microstructure. A build-from-scratch module consolidates practice: data preparation, model selection, and training/testing pipelines, revisiting crystal hardness in a second exercise. The course then turns to accelerated discovery via robotics and self-driving laboratories, Bayesian optimisation, and reinforcement learning, explored through a closed-loop optimisation exercise. We conclude with generative AI—large language models, transitions from latent spaces to diffusion models, and agentic research workflows—culminating in a capstone research challenge.

Artificial Intelligence in materials science, Machine learning, Materials data & FAIR principles, Crystal representations (composition–structure–property), Generative Artificial Intelligence (LLMs, diffusion models)

03-09-2025