Tutorial MT02—Large Language Models and Foundation Models for Materials Science

Instructors: Jason Hattrick-Simpers, University of Toronto; Indra Priyadarsini, IBM Research–Tokyo; Seiji Takeda, IBM Research–Tokyo

The emergence of open-source foundation models in materials science is revolutionizing how researchers approach molecule design, property prediction and materials discovery. This hands-on tutorial is designed to introduce participants to these powerful tools, focusing on SMILES, SELFIES and molecular graph-based models. Participants will gain practical experience in fine-tuning these models with custom data sets, visualizing molecular embeddings and executing downstream tasks such as property prediction and molecule generation. Through interactive sessions, attendees will learn how to leverage these models to accelerate their R&D processes. The tutorial will walk participants through the entire workflow, from data preprocessing and model customization to evaluating results and interpreting insights.

Learning Objectives include:

• Understand SMILES, SELFIES and graph-based encodings for molecules, including how each approach impacts model training and prediction accuracy
• Learn how to adapt pretrained open-source models to specific materials data sets, enabling more accurate predictions and property modeling
• Use visualization tools to interpret embeddings and understand how models represent different molecules, aiding in feature interpretation and insights into chemical space
• Utilize fine-tuned models in real-world applications like predicting materials properties or generating novel molecular structures with targeted properties
• Incorporate foundation models into existing computational pipelines and best practices for combining model predictions with experimental planning and validation

Tutorial Schedule

1:30 pm

Introduction to Foundation Models in Materials Science, 

Seiji Takeda, IBM Research–Tokyo, Japan

2:00 pm

Exploring SMILES, SELFIES and Graph Models with WebUI

Seiji Takeda, IBM Research-Tokyo, Japan; Indra Priyadarsini, IBM Research–Tokyo, Japan

2:45 pm BREAK

3:15 pm

Hands-On Coding with SMILES, SELFIES and Graph Models in Python

Indra Priyadarsini, IBM Research–Tokyo, Japan

4:00 pm

Practical Model Fine-Tuning for Real-World Applications

Indra Priyadarsini, IBM Research–Tokyo, Japan; Seiji Takeda, IBM Research–Tokyo, Japan; Jason Hattrick-Simpers, University of Toronto, Canada