DuPont, Microsoft and NobleAI Team Up on AI-Driven R&D

Holographic molecular structures floating above a materials science lab bench
AI-generated illustration

Materials science is slow because every promising idea eventually needs a lab test, and labs are expensive. On November 12, 2026, DuPont said it is working with Microsoft and NobleAI to move more of that exploration into software, in a collaboration it says is designed to reduce cycle times for delivering customer solutions, per the company’s press release.

Three partners, three roles

DuPont (NYSE: DD) describes the collaboration as combining its application expertise, NobleAI’s Science-Based AI modeling, powered by the Visualizations, Insights and Predictions (VIP) Platform, and Microsoft Discovery, which Microsoft calls its enterprise agentic AI platform for R&D. The stated aim is to help DuPont scientists and engineers evaluate potential solutions more rapidly, guide experimentation and speed up cycles of learning.

In practice the division of labor is straightforward. Microsoft Discovery is the layer where scientists plan and orchestrate research. NobleAI supplies the scientific models that predict and optimize. DuPont brings the data, the domain knowledge and the customer problems.

What science-based AI means here

NobleAI says its approach integrates scientific principles throughout the AI modeling workflow. Its company description in the release says the VIP Platform turns limited industrial data into predictive models that organizations can act on, so they can evaluate more options in software before committing to physical testing. That matters in chemistry and materials, where data sets are often small and expensive to produce, and a purely data-hungry model may have too little to learn from.

NobleAI CEO Alex Wang said the combination lets DuPont’s teams build on the data and expertise they already have, within the workflows they already use, to explore far more ideas in silico, through computer simulation, so physical testing goes to the candidates most worth pursuing. NobleAI is also described as one of the first launch partners for Microsoft Discovery, and it is backed by M12 and Chevron Technology Ventures.

Governance is part of the pitch

DuPont Chief Information Officer Matt Abbott framed the work around trust. He said the collaboration reflects how DuPont is scaling AI with security and governance designed in from the start, so what proves out in one lab can be put to work across the company. DuPont Chief Technology Officer Marty DeGroot said the goal is to expand access to digital innovation tools that help teams accelerate development, shorten time to market and extend the impact of innovation across the portfolio.

Microsoft’s Aseem Datar, Corporate Vice President of Product Innovation for Microsoft Discovery and Quantum, said the platform connects R&D teams’ data, models and know-how so they can explore more possibilities with confidence and rigor.

What it means

The announcement is light on numbers. The release does not give a target for time savings, a list of products affected or a rollout date, so it is best read as a statement of direction rather than a measured result. Even so, it points to patterns worth watching:

  • R&D is becoming an agentic workflow. Platforms that plan experiments and call specialist models are moving from demos into established industrial companies.
  • Small data needs physics. Models that embed scientific principles may suit industries where experiments are costly and records are limited.
  • Simulate first, test second. The value case is fewer wasted lab runs, not the removal of physical testing.
  • Governance scales with ambition. Enterprises want proof in one lab to be repeatable across the company without losing control of data.

The bottom line

DuPont is putting a science-first AI stack and an agentic R&D platform behind its scientists, with the goal of shorter development cycles. The proof will be in the outcomes DuPont reports later, such as faster customer solutions, which the company has not yet quantified.

Sources