AI4X Tutorials

Abstract: Advanced materials are essential for solving critical global challenges such as energy, sustainability, healthcare, and resource scarcity. Traditional discovery methods, however, remain slow, costly, and inefficient. This tutorial introduces theoretical frameworks and methodologies behind combinatorial material studies, exploring how data-driven techniques and automated atomic manufacturing can dramatically accelerate materials discovery. Participants will learn how combinatorial approaches—such as rapid prototyping, high-throughput screening, and accelerated design of experiment—can enable faster identification and optimization of functional materials, paving the way for transformative innovations across diverse fields.

Goals:

• Educate attendees on combinatorial material discovery methodologies.
• Showcase theoretical concepts behind screen and optimization of materials using atomic precision techniques.
• Highlight the potential and current limitations of combinatorial approaches.
• Foster discussion on collaboration, especially integrating AI techniques.

Abstract: Experimental research in chemistry, biology, and materials science is often limited by slow, manual workflows that demand significant time, expertise, and physical resources. Self-driving laboratories offer a powerful solution by combining automation with artificial intelligence to accelerate discovery and improve reproducibility. This tutorial provides a practical introduction to the core principles behind self-driving labs, aimed at researchers who may not have prior experience with laboratory automation, programming, or statistical modelling. The tutorial is structured into four sections. The first introduces laboratory automation, focusing on how to design an automated experimental system, select appropriate devices, and establish communication between components. The second section explores Bayesian optimisation as a method for guiding experimental decisions, with a focus on intuitive understanding rather than formal theory. The third section demonstrates how automation and optimisation can be integrated to create a fully autonomous self-optimising flow reactor for chemical synthesis. The final section includes a live demonstration using AMLearn, a platform developed by the presenter, to illustrate how the ideas discussed can be implemented in practice. Emphasis is placed on general principles rather than specific tools, enabling participants to begin applying self-driving methods within their own research environments. The tutorial aims to lower the entry barrier to this emerging area of research and support wider adoption of autonomous experimentation across the natural sciences.

Goals:

• Understand the core hardware components required to build a self-driving laboratory and how to coordinate them effectively.
• Recognise the communication protocols and control strategies used to integrate laboratory devices into a single automated system.
• Gain an intuitive understanding of Bayesian optimisation, including surrogate models and acquisition functions, without requiring prior statistical training.
• Learn how to connect automation and machine learning into a closed-loop experimental workflow.
• Be able to identify opportunities to apply self-driving methods using existing equipment in their own laboratories, including standard tools not typically associated with automation.
• Develop the confidence to evaluate, adapt, or begin designing autonomous experimentation systems within their own research environments.

Abstract: This tutorial will explore how artificial intelligence (AI), particularly Graph Neural Networks (GNNs), is transforming scientific workflows for complex materials research. Attendees will gain practical insights into automating computational experiments, training GNNs on real-world datasets, and accelerating materials discovery. The session will include hands-on exercises, equipping researchers with the skills to integrate AI-enhanced scientific workflows into various domains, including chemistry and biology. This tutorial aims to provide a balanced and insightful exploration of AI-driven workflows in materials science, with a focus on reproducibility and accessibility for the AI4X community.
A) Theoretical Foundations of Ab Initio Methods
B) Classical Workflow Construction for Materials
C) Theoretical Foundations of Graph Neural Networks
D) Machine Learning-Enhanced Workflow Design
E) Hyperparameter Experimentation and Performance Optimization
F) Hands-on Session and Reproducibility

Goals:

• Introduce participants to AI-driven scientific workflows and their applications in materials science.
• Provide hands-on learning with workflow automation tools.
• Showcase how GNNs enhance predictions of complex material properties.
• Highlight computational efficiency gains through AI integration.

Abstract: Climate change poses an existential threat to humanity, with the growing climate energy imbalance driving more frequent and severe extreme weather events. Accurate forecasting of such events is crucial for timely action but remains computationally expensive using traditional methods. Recent advances in AI Numerical Weather Prediction (AINWP) models offer a transformative approach - delivering forecasts thousands of times faster than numerical simulations - enabling large ensemble predictions for extreme events. However, identifying the precursors to these extremes, and understanding how they evolve under climate change, remains an open challenge. In this tutorial, we will first demonstrate how AINWP models can predict the track and intensity of a tropical cyclone. We will then introduce an interpretable machine learning framework, XAI4Extremes, which generates relevance weather maps to highlight key extreme-weather precursors identified by deep learning models.

Goals:

• Showcase the practical application of AI models for accurate and timely predictions of real-world extreme weather events.
• Attendees will learn how to train and run a pre-trained model to forecast actual events, and how to use these models for generating diagnostic results.
• Gain hands-on experience with the XAI4Extremes framework to interpret deep learning outputs and understand the key weather precursors identified by AI.
• Compare machine-generated interpretations with domain knowledge to enhance understanding of extreme-weather mechanisms, including how these mechanisms evolve over multi-year periods due to climate change.

Abstract: This tutorial presents an overview of how large language models (LLMs) are reshaping the landscape of scientific discovery. We cover critical advancements in applying LLMs to fundamental tasks in the scientific discovery process, including: (1) initial hypothesis direction search, (2) fine-grained hypothesis formulation from coarse-grained hypothesis direction, (3) interaction between experimental results and hypothesis refinement, (4) automatic novelty evaluation of discovered hypotheses, and (5) benchmarking LLMs for scientific discovery. Participants will gain a bird’s-eye view of the current field, as well as in-depth insights into each of these directions. Hands-on experience will also be provided through interactive demos showcasing LLMs for scientific discovery.

Goals:

• Understand the emerging role of LLMs in automating key tasks in scientific discovery, including hypothesis generation, refinement, and evaluation.
• Learn how to mathematically formulate scientific discovery tasks and apply LLM-based agentic frameworks to solve them.
• Analyze current challenges, solutions, and research findings in using LLMs for scientific reasoning.
• Gain practical experience through interactive demos involving hypothesis generation, refinement, and automatic novelty evaluation.
• Reflect on open challenges, ethical considerations, and future directions for benchmarking and advancing LLM-driven scientific discovery.