Materials science is entering a new era where artificial intelligence doesn't just predict material properties, but actively designs and validates new candidates autonomously. MIT Professional Education has launched a four-day intensive course that teaches engineers and scientists how to build AI agents capable of scientific reasoning, moving the field from traditional data-driven prediction to what experts call "agentic AI".

What Is Agentic AI in Materials Discovery?

Agentic AI represents a fundamental shift in how artificial intelligence approaches materials science. Rather than simply analyzing data and making predictions, these systems can plan, execute, and refine scientific tasks independently. They can read scientific literature, formulate hypotheses, write and execute code, run simulations, and even interface with experimental automation pipelines to suggest validation experiments.

The course, titled "Applied AI for Materials Discovery," was held July 27-30, 2026, and was led by MIT Professor Markus J. Buehler. The program emphasizes moving beyond static predictive models to dynamic agents that can reason about physics, chemistry, and biology in ways that mirror how human scientists think.

"In this course, you won't just watch AI in action, you'll collaborate with it, building agents that reason, design, and solve problems alongside you, for you, and teaching you," explained Professor Markus J. Buehler.

Markus J. Buehler, Professor at MIT

How Are Foundation Models Changing Materials Design?

Foundation models, the large AI systems trained on vast amounts of text and data, are being adapted to understand materials science in new ways. The MIT course teaches participants how to harness multimodal foundation models that can integrate text, images, graphs, and three-dimensional voxel data into a unified reasoning framework. This allows AI systems to understand the full context of a material and support retrieval, prediction, and design tasks simultaneously.

One of the critical innovations is bridging what researchers call the "reality gap." This means connecting AI-generated designs directly to physical constraints using multiscale modeling, ensuring that the materials AI proposes are actually manufacturable and viable in the real world. The course emphasizes explicit physical constraints and verification loops through simulation and, where applicable, experimental validation.

Steps to Building Enterprise-Grade AI Discovery Workflows

• Orchestrate Agentic Workflows: Design systems that can read scientific literature, formulate hypotheses, write and execute code, run simulations, and interface with experimental automation to suggest testable validation plans.
• Apply Multimodal Foundation Models: Integrate text, images, graphs, spectra, and three-dimensional representations to reason over materials context end-to-end and perform inverse design tasks.
• Bridge Physics Constraints: Couple learning to multiscale modeling from atomistic to continuum scales, ensuring AI-proposed candidates are grounded in real physical laws and can be manufactured.
• Convert Unstructured Knowledge: Use vision-language and document understanding models to transform lab notebooks, legacy PDFs, reports, and micrographs into structured, actionable insights.
• Implement Governance Practices: Establish interpretability and traceability mechanisms suitable for high-stakes engineering decisions, including uncertainty quantification and adversarial evaluation.

The curriculum covers state-of-the-art generative models, including diffusion and flow-based methods, as well as graph generative approaches that can propose novel material candidates under explicit constraints. Participants also learn to use modern surrogate physics models, such as neural interatomic potentials and physics-informed neural networks (PINNs), to accelerate simulation-informed discovery.

New for 2026, the course introduces massively parallelized AI agents and swarm intelligence that can solve complex problems collaboratively. The curriculum also emphasizes the emerging category of "AI scientist," systems that can ingest vast amounts of unstructured data from handwritten lab notebooks to legacy PDFs and structure it into actionable insights automatically, unlocking decades of dormant value in organizations.

"By learning the fundamentals of AI, you can make informed decisions about which models and agents to use, and how to apply them strategically, avoiding the common pitfall of rushing into technology without understanding it," noted Professor Markus J. Buehler.

Markus J. Buehler, Professor at MIT

The course cost $3,600 and was delivered in a live online format, allowing participants from around the world to engage in real-time coding exercises. Participants received dozens of code examples and datasets they could immediately apply to their own projects. The program awarded 2.2 continuing education units (CEUs) and a Certificate of Completion from MIT Professional Education.

Beyond lectures, participants left the course with practical assets including ready-to-use agent templates, customizable Jupyter notebooks covering the entire pipeline from data curation to physics verification, and datasets for hands-on experimentation. These tools are designed to be immediately deployable in organizational materials discovery workflows.

The shift toward agentic AI in materials science reflects a broader transformation in how artificial intelligence is being applied to scientific research. Rather than replacing human scientists, these systems are designed to augment human expertise, accelerating the discovery process and enabling researchers to explore a vastly larger design space than would be possible through traditional methods alone.