MIT Short Course: Machine Learning for Materials Informatics (Jul 29 - Aug 2, 2024)

Dera iMechanica Community,

Below is information about a short course I will be offering at MIT this summer, in Live Virtual format: Machine Learning for Materials Informatics (Jul 29 - Aug 2, 2024).  This is an exciting opportunity that will cover fundamentals and applications in the emerging space of AI/ML for engineering, featuring hands-on interactive code development in Jupyter notebooks. We'll do a deep dive into all critical tools from autoencoders to graph neural nets to multimodal LLMs and multi-agent modeling. Please reach out to me if you have any questions.

We have a few special fellowships for postdocs and students that cover part of the course fee. Please reach out to me with a brief CV for consideration. 

Markus Buehler
McAfee Professor of Engineering, MIT

Machine Learning for Materials Informatics

Instructor: Prof. Markus J Buehler, mbuehler@MIT.EDU

Jul 29 - Aug 2, 2024 (3.5 days)

Learn more here: https://professional.mit.edu/course-catalog/machine-learning-materials-i...

With the emergence of physics-based generalizable deep learning tools, the materials community is at the verge of unprecedented breakthroughs, leveraging materials modeling, analysis, and design toward a more efficient, less costly, and more versatile response to material property demands, as well as research and economic opportunities. This course will teach you how to tap into your existing data and develop an actionable vision for incorporating material informatics into current research strategies for developing technologies, services and new investigation directions. Moreover, with data available from autonomous experimentation or large databases like the Materials Genome initiative, there exist many opportunities to accelerate your materiomic design platform.

In this course you will fully learn how to incorporate these new technologies and methods into your own material modeling, analysis and design processes in order to capitalize on recent AI breakthroughs, such as multimodal large language models or LLMs that have had significance impact over the past year (e.g. GPT-4, GPT-4o/ChatGPT, Gemini, etc.), open-source models like Llama-3, Phi-3 or Mistral, DNA and protein models (e.g., AlphaFold), diffusion models for molecular, protein and microstructure design, graph neural networks applied from molecular to macroscale structures, and a host of other methods adapted specifically for the analysis, design and modeling of materials. An exciting field we will cover is multi-agent AI modeling that brings together multiple systems to solve complex science and engineering problems, autonomously. You will also learn how to develop, train and validate your own custom models - either derived from pretrained foundation models, or building them from scratch. 

The course involves a mix of lectures, hands-on labs and clinics, custom data and generative analyses, for an immersive experience. Participants will learn fundamentals and hands-on techniques to deploy machine learning in materials development and gain first-hand understanding of state-of-the art tools for varied applications ranging from data mining to inverse design. We will cover scales from the molecular to the continuum and you will develop a deep understanding of machine learning especially in the context of engineering problems. The course includes various case studies, including a deep dive into the use of OpenAI and other APIs for engineering problems, fine-tuning custom vision LLMs, efficient large model serving using Rust, and working with a variety of open-source language models, the Hugging Face ecosystem, and more.  You will specifically learn how to fine-tune large multi-modal foundation models to solve your own tasks in a variety of engineering settings.

Specific topics covered:

• Modern machine learning tools, especially focused on deep learning (includes: convolutional neural nets, adversarial methods, graph neural nets, transformer and language models, diffusion models; neural molecular dynamics; Bayesian optimization; autonomous self-driving labs; fine-tuning of LLMs and related foundation models, multi-modal LLMs and multi-agent modeling for scientific problem solving and discovery)
• Analysis of images, voxel data, dynamical data, and graphs, as well as language and symbolic methods and hybrid and multimodal approaches, including graph reasoning strategies 
• Visualization and data analysis methods, including statistical methods, graphic rendering, virtual reality, multi-modal data analysis and pattern detection
• Data mining and dataset construction, especially focused on building datasets from complex multimodal sources (text, papers, patents, images, video, etc.) and extracting critical insights or training custom models for use in multi-agent AI systems

The instructor will masterfully break down this complex field into easy-to-digest concepts, to offer you direct access to leverage the new tools for your problem space, and to develop the skill to judge and assess the best tools for the job. Alongside peers from around the world, you will engage in interactive lectures and hands-on coding clinics and labs delivered in a live virtual format. These activities are designed to help you learn, design, and apply modern material informatics tools—specifically artificial intelligence and machine learning—including neural interatomic potentials, large-scale multiscale modeling to improve the speed, efficiency, and cost effectiveness of your discovery, prototyping, and development processes. You will learn how modern computational tools enable us achieve almost any desirable accuracy in multiscale material discovery, connecting quantum to the macro-world. 

https://professional.mit.edu/course-catalog/machine-learning-materials-i...