A Researcher's Guide to AI in Materials Science: From First Steps to a Frontier

Thank you to everyone who attended my tutorial at the McMaster CCEM Summer School. The energy and insightful questions were truly inspiring. As promised, this post serves as a detailed guide for anyone looking to begin their journey into the exciting intersection of artificial intelligence, machine learning, and materials science.

The central message of my talk was that AI is not a replacement for rigorous scientific inquiry, but a powerful amplifier for it. Our goal is to teach our microscopes not just to see, but to understand. This is a transformative endeavor, and while it may seem daunting, the path to getting started is more accessible than ever.

This guide provides a curated list of resources to take you from foundational concepts to the cutting edge of research.

1. Building a Conceptual Foundation (Online Courses)

Before diving into complex code or research papers, it's crucial to build a strong conceptual intuition for how machine learning works. These courses are exceptional starting points that prioritize understanding over abstract mathematics.

• Machine Learning Specialization by Andrew Ng (Coursera)

  • Who it's for: The absolute beginner. This is widely considered the definitive "first course" in machine learning for a reason.

  • What you'll learn: Dr. Ng has a gift for explaining the core ideas of supervised and unsupervised learning, including regression, classification, and clustering, with brilliant clarity. You will come away understanding the "why" behind the algorithms.

• Deep Learning Specialization by Andrew Ng (Coursera)

  • Who it's for: Those who have completed the introductory course and want to understand the architecture of modern AI.

  • What you'll learn: This specialization demystifies neural networks, convolutional neural networks (CNNs) for image analysis, and the strategies used to train these powerful models effectively.

• Practical Deep Learning for Coders by fast.ai

  • Who it's for: Those who prefer a "top-down," code-first approach. It’s perfect for the experimentalist who wants to start applying models quickly and learn the theory along the way.

  • What you'll learn: You'll learn how to build and train state-of-the-art deep learning models using the PyTorch library with a focus on practical results and best practices.

2. The Practitioner's Toolkit (Key Software & Libraries)

Your research will ultimately be done through code. The open-source community has built an incredible ecosystem of tools that are free to use and supported by excellent documentation. All these tools are primarily based in the Python programming language.

• The Deep Learning Frameworks: TensorFlow & PyTorch

  • These are the two titans of deep learning. You don't need to know both; choose one and go deep. They provide the fundamental building blocks for creating and training neural networks. My group has used both to great effect in our work.

• The Classical ML Workhorse: Scikit-learn

  • For any task that doesn't require a deep neural network (e.g., clustering, classical regression, dimensionality reduction like PCA), Scikit-learn is the gold standard. It’s incredibly user-friendly and powerful.

• Model Exchanges: Huggingface

  • There are many models that have been developed that you can adapt for your own problems. Huggingface is one such exchange that can help you quickly get up to speed.

• Domain-Specific Tools for Microscopy & Materials:

  • Atomap: A foundational open-source tool for finding and fitting atomic column positions in high-resolution images using 2D Gaussian fitting. Our TEMWizard software was built to provide a graphical interface for this library.

  • AtomAI: A powerful PyTorch-based toolkit that connects deep learning models to experimental microscopy data, enabling analysis of atomic and mesoscale images and spectroscopy.

  • HyperSpy: An essential open-source Python library for analyzing multi-dimensional data, including electron energy loss spectroscopy (EELS) and energy-dispersive X-ray spectroscopy (EDS) datacubes.

  • py4DSTEM: A specialized toolkit for the analysis of 4D-STEM datasets, a rapidly growing area of research.

3. Seminal & Foundational Papers: Your Reading List

Staying current means reading the literature. This list is not exhaustive, but it provides a starting point with papers that established a new concept, provide a vision for the future, or demonstrate a particularly insightful application in our field. I have organized them by theme to help guide your reading.

The Deep Learning Revolution (General)

• Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). "ImageNet classification with deep convolutional neural networks." Communications of the ACM, 60(6), 84-90.

  • Why it’s important: This is the "AlexNet" paper that kickstarted the modern deep learning revolution. It demonstrated that deep neural networks, trained on massive datasets with GPU hardware, could dramatically outperform all previous methods in image classification.

Core Task: Feature Identification and Segmentation

• Nord, M., et al. (2017). "Atomap: a new software tool for the automated analysis of atomic resolution images using two-dimensional Gaussian fitting." Advanced Structural and Chemical Imaging, 3(1), 9.

  • Why it’s important: This paper details the classic model-based approach to finding atoms. It is a powerful technique and provides an important baseline for understanding the advantages of deep learning-based methods.

• Ziatdinov, M., et al. (2017). "Deep Learning of Atomically Resolved Scanning Transmission Electron Microscopy Images: Chemical Identification and Tracking Local Transformations."ACS Nano, 11(12), 12742-12752.

  • Why it’s important: This was an important paper demonstrating how deep learning could be used to find and classify atomic-scale features and defects directly from experimental STEM data, paving the way for many of the applications we see today.

• Akers, S., Kautz, E., et al. & Spurgeon, S.R. (2021). "Rapid and flexible segmentation of electron microscopy data using few-shot machine learning." npj Computational Materials, 7(1), 187.

  • Why it’s important: This work from my team addresses a key bottleneck in real-world science: the "sparse data" problem. We show how few-shot learning can be used to segment complex microstructures with minimal human-labeled examples, dramatically speeding up analysis.

Advanced Task: Discovery, Forecasting, and New Modalities

• Ter-Petrosyan, A.H., et al. & Spurgeon, S.R. (2023). "Unsupervised segmentation of irradiation-induced order-disorder phase transitions in electron microscopy." Proceedings of the Thirty-Seventh Conference on Neural Information Processing Systems (NeurIPS).

  • Why it’s important: This paper demonstrates a true discovery capability. We developed an unsupervised model that can identify hidden order-disorder transitions in irradiated materials without any prior human labels, showcasing the power of AI for exploring unknown phenomena.

• Lewis, N., Jin, Y., et al. & Spurgeon, S.R. (2022). "Forecasting of in situ electron energy loss spectroscopy." npj Computational Materials, 8(1), 252.

  • Why it’s important: Moving beyond static images, this work shows how recurrent neural networks can forecast the evolution of spectroscopic signals during in situ experiments. This is a critical step toward predictive control of dynamic processes at the nanoscale.

• Liu, Y. et al (2023). "Learning the right channel in multimodal imaging: automated experiment in piezoresponse force microscopy." npj Computational Materials 9, 34.

  • Why it’s important: This work shows the breadth of ML applications beyond TEM/STEM, demonstrating a sophisticated deep kernel learning approach to automatically analyze complex domain structures in piezoresponse force microscopy (PFM) data.

The Vision: Reviews, Perspectives, and Autonomous Science

• Spurgeon, S.R., Ophus, C., et al. (2021). "Towards data-driven next-generation transmission electron microscopy." Nature Materials, 20(3), 274-279.

  • Why it’s important: My collaborators and I laid out a community perspective and a vision for the future of microscopy, arguing for a move toward data-centric, AI-driven workflows to accelerate discovery.

• Olszta, M., et al. & Spurgeon, S.R. (2022). "An automated scanning transmission electron microscope guided by sparse data analytics." Microscopy and Microanalysis, 28(5), 1611-1621.

  • Why it’s important: Here, we moved from post-analysis to active experimentation. We detailed the development of the "AutoEM" platform, which embeds AI directly into the microscope control loop, enabling the instrument to make intelligent decisions on the fly.

• Dyck, O., Ziatdinov, M., Lingerfelt, D.B. et al. (2019). "Atom-by-atom fabrication with electron beams." Nat Rev Mater 4, 497–507.

  • Why it’s important: This paper reviews the incredible progress at ORNL autonomous manipulation of individual atoms to create custom atomic structures based on a predefined blueprint.

4. Communities & Research Groups to Follow

Science is a collaborative effort. Following the work of leading groups and participating in community events is a great way to stay inspired and informed.

• Influential Research Groups:

  • Sergei V. Kalinin (University of Tennessee, Knoxville): A true pioneer in the application of AI and automated experimentation to scanning probe and electron microscopy.

  • Alán Aspuru-Guzik (University of Toronto): A leader in the broader concept of "self-driving laboratories" for accelerating chemical and materials discovery.

  • My Research Group and Collaborators (NREL / Colorado School of Mines): My work focuses on autonomous science for materials in energy and extreme environments.

• Professional Societies & Meetings:

  • Microscopy Society of America (MSA): The annual Microscopy & Microanalysis meeting is the premier venue for our field. We regularly organize symposia on AI and data science there.

  • Materials Research Society (MRS): The MRS Spring and Fall meetings are essential for materials scientists, with a rapidly growing number of symposia dedicated to data science, AI, and automated discovery.

I hope this guide provides you with a solid starting point and a clear roadmap. This is a journey of continuous learning, but one that holds immense promise for the future of materials science. The tools are here, the community is growing, and the scientific frontiers are waiting.

I look forward to seeing what you discover!