Session chairs: Prof. Siddhartha Mishra (Computational and Applied Mathematics Laboratory, ETH Zurich), Prof. Bernd Bickel (Computational Design, ETH Zurich)

Session organizers: Karin Yu (ETH AI Center), Dr. Fanny Lehmann (ETH AI Center)

AI is driving transformative progress across the natural sciences and engineering. By enabling the discovery of underlying physical laws directly from data, accelerating simulations through surrogate modeling, and integrating data-driven with physics-based approaches, AI is reshaping how scientific knowledge is generated and applied. These advances are unlocking new opportunities, but they also raise critical questions about the design of appropriate AI architectures, the availability and quality of training data—particularly in domains where data collection is expensive or experimentally constrained—highlighting the need for robust, interpretable, and generalizable models.

 

This session will explore recent advances at the intersection of AI and science, with applications spanning biology, fluid dynamics, engineering, neuroscience, and beyond. We will discuss methodological innovations, including physics-informed machine learning, generative models, and uncertainty quantification. A panel discussion will address the challenges and limitations of current AI techniques and conclude with perspectives on how AI can be harnessed to reduce engineering costs, accelerate innovation, and contribute to more sustainable design and decision-making.

 

 

Prof. Elizabeth Cross (Univ. of Sheffield) A spectrum of physics-informed machine learning approaches for problems in structural dynamics

As monitoring data from our critical systems and structures become more abundant, engineers (should) naturally wish to benefit from the learning available from them. Indeed, many elements of structural assessment and, in particular, those relying on a dynamic signature, are now evolving to take advantage of this, leading to the creation and adoption of a wealth of data-driven approaches. The use of machine learning in structural health monitoring, for example, is common, as many of the inherent tasks (such as regression and classification) in developing condition-based assessment fall naturally into its remit.

A significant challenge here, that is not often acknowledged, however, is that we commonly lack representative data from across the range of environmental and operational conditions structures will undergo, limiting the usability of an entirely data-based approach.

This talk will present a number of ways of incorporating the physical insight an engineer will often have of the structure they are attempting to model or assess into a machine learning approach through a Gaussian process regression framework. The talk will demonstrate how grey-box models, that combine simple physics-based models with data-driven ones, can improve predictive capability for structural assessment and system identification tasks. A particular strength of the approaches demonstrated here is the capacity of the models to generalise, with enhanced predictive capability in different regimes, increasing applicability in light of the aforementioned challenge.

 

Prof. Nils Thürey (Technical University of Munich) Generative AI & Probabilistic Learned Solvers

In this talk, I'll explore the potential of generative AI techniques (such as denoising diffusion and flow matching) for learned simulators. These diffusion models transform learned representations into probabilistic models, enabling the generation of samples from a posterior distribution rather than merely producing a deterministic estimate of the mean. Furthermore, integrating existing numerical methods into learning tasks allows for the creation of highly accurate inverse solvers. The ability of learned probabilistic surrogates to sample from the posterior is particularly promising for challenging downstream tasks, such as uncertainty quantification and improved interpretability.

 

Prof. Basile Wicky (ETH Zurich) Interfacing with biology using protein design

Recent advances in de novo protein design are transforming biology into an increasingly generative discipline. Fueled by progress in machine learning, we can now generate novel proteins in silico with increasing precision—expanding the design space beyond what evolution has explored. In the first part of my talk, I will describe how protein structure prediction networks, originally built for inference, can be re-purposed as generative models through strategies like hallucination.

Beyond generating isolated proteins, we are now exploring how to compose them into functional systems. I will share recent work where we use designed protein networks to implement Boolean logic directly in mammalian cells. Looking forward, we envision extending this approach toward in situ cellular classification at the molecular level—treating designed proteins as modular building blocks for biological computation. These developments point to the potential of generative protein models as a new way to interact with and program biological systems.

 

Dr. Michał Januszewski (Google Research) Mapping the structure and function of the brain with AI

To understand how the brain's function arises from its physical wiring, we must first map its structure. A single cubic millimeter of brain tissue generates over a petabyte of image data, creating an immense data challenge that makes AI essential for analysis at scale. This talk will demonstrate how we use AI to automatically reconstruct neurons and their connections from massive volumetric microscopy datasets. We will then explore how these structural maps are used to build computational models that directly link the brain's anatomical structure to its dynamic function.