Description
Full title: How to use CrystaLLM-pi to predict crystal structures with desired functional properties, or matching experimental data and adapt the model to your own data
In this tutorial, we will cover how the user can load up and make predictions with pre-packaged deep learning models to generate materials structure with a target property or matching a measured experimental data signal on their own devices. Additionally, we will show how to format your own data in order to adapt the model so it can make structure predictions conditioned on a desired property.
Learning outcomes
- Increased understanding on the inside workings of a specialized transformer model that can generate crystal structures conditioned on functional properties
- Gain the skills to load pre-trained models and generate materials with desired attributes on your own device
- Understand how to format structure and property data so it can be used for training a transformer
- Learn how to adapt a toy model to this dataset in order to understand how it could be applied to your data
Audience
Computational materials scientists
Pre-requisite background
GitHub usage to access and download software; Familiarity with Python and code execution in computational notebooks, e.g. Jupyter Notebook; Python environment basic understanding.
Pre-requisite setup
A Hugging-Face API key. A markdown file with instructions on how to do this is provided.
Tutorial outline
Background
- Crystallographic Information File
- Transformers/LLMs for materials generation
- Conditioning on Functional Properties
Basic usage
- Loading and generating desired materials using pre-trained publicly available CrystaLLM models
Fine-tuning
- Make a data frame containing CIFs and properties of interest in correct ML training format
- How to link up your dataset to CrystaLLM to fine-tune it yourself
- How to choose the correct parameters to train your model
- Train your model with a toy example
- Load a state of the art trained model & generate structures with it