AI-driven Scientific Discovery

The recent success of generative Artificial Intelligence (AI) has shown how powerful data-driven models can be for creating images and text. Still, industrial applications require AI systems that support engineering intelligence rather than content creation. The research in the business area focuses on scientific AI methods that integrate (potentially small) experimental data, expert knowledge (for example, expressed through preferences), and physics-based simulations in order to build learning systems that can solve many design problems.

An important paradigm is inverse design, in which learning systems come up with designs or parameters that achieve desired target behaviors or properties. Beyond optimization, these learning systems systematically explore trade-offs and identify the limits of what is achievable, helping industry reduce trial-and-error, focus experiments, and make informed design decisions under real-world constraints.

The application domains span a wide range of scales and modalities, including inverse design of molecules and materials, as well as data-driven discovery of (meta)materials under physical and manufacturing constraints.

• Post Hoc Regression Refinement via Pairwise Rankings
• Trustworthy Inverse Molecular Design via Alignment with Molecular Dynamics
• Autoinverse: Uncertainty Aware Inversion of Neural Networks
• Shape from Release: Inverse Design and Fabrication of Controlled Release Structures
• Topology Optimization via Frequency Tuning of Neural Design Representations