|
9-9 Mar 2026 Paris (France)
|
|
Description Summary of the event (see below for the Programme and presentation slides) AI4Science at Sorbonne Université: toward “centaur science” On March 9th, Sorbonne Université hosted the launch of the AI4Science initiative at SCAI (Sorbonne Cluster for Artificial Intelligence), bringing together researchers from multiple disciplines, institutional representatives, and industrial partners to explore how artificial intelligence is reshaping scientific discovery. Beyond a scientific showcase, the event highlighted a deeper transformation: the emergence of a new research paradigm in which AI and scientific reasoning are increasingly intertwined. This transformation was framed by two complementary keynote talks. Karthik Duraisamy (University of Michigan) emphasized the rise of scientific foundation models and the emergence of AI-driven discovery engines: systems in which large-scale models, combined with agentic AI, can autonomously explore hypotheses, design experiments, and accelerate scientific workflows. In parallel, Jesse Thaler (MIT) introduced the concept of “centaur science”, a hybrid paradigm combining artificial intelligence and domain expertise. Advancing science in this context requires not only teaching machines to “think like physicists”, by embedding symmetries, constraints, and robustness, but also encouraging scientists to “think like machines”, embracing data-driven reasoning and algorithmic thinking. Together, these perspectives point toward a profound redefinition of the scientific process. Across disciplines, the scientific sessions illustrated how AI is moving beyond mere acceleration toward genuine discovery. In chemistry, Jean-Philip Piquemal (LCT) presented FeNNix-Bio1, a quantum-based foundation model for atomistic simulations in drug design, capable of capturing complex molecular interactions with high accuracy while scaling to large systems, bridging the gap between quantum chemistry and molecular modeling. In materials science, Marco Saitta (LPENS) highlighted how AI-driven simulations and machine-learned interatomic potentials are accelerating the discovery of materials for the energy transition and illustrated how the research ecosystem in AI and materials science is evolving thanks to large scale consortia and initiatives. In biology, Alessandra Carbone (LCQB) showed how deep learning has transformed our understanding of the protein universe, from structure prediction to system-level functional inference, enabling large-scale reconstruction of interaction networks. In climate science, Claire Monteleoni (INRIA-ARCHES) presented generative AI approaches for forecasting and downscaling, achieving high performance with significantly reduced computational cost. In astrophysics, Guilhem Lavaux (IAP) demonstrated how combining Bayesian inference and machine learning makes it possible to extract new physical insights from observational data, even in the absence of controlled experiments. Finally, Patrick Gallinari (ISIR) discussed advances in physics-aware AI, from neural operators to foundation models for dynamical systems, emphasizing the importance of physical consistency, generalization, and hybrid modeling. A strong convergence emerges from these contributions. First, AI is increasingly used not only to predict but to discover, revealing hidden mechanisms in complex systems. Second, the dominant paradigm is becoming hybrid, combining data-driven models with physical knowledge, simulations, and statistical inference. Third, many domains are moving toward foundation models for science, designed to generalize across systems, scales, and tasks. These scientific advances raise important strategic and societal questions, which were at the heart of the two round tables bringing together leading actors from academia and industry, including A. Duval (Entalpic), R. Elie (DeepMind), S. Gigan (LightOn), J.-P. Piquemal (Sorbonne Université and Qubit Pharmaceuticals), G. Biau (Sorbonne Université–SCAI), A. Butter (LPNHE, IML-CERN), K. Duraisamy (University of Michigan), J. Fadili (CNRS-AISSAI), M. Husson (CNRS-LTE), and J. Thaler (MIT), with discussions moderated by P. Gallinari and P. Cinnella. Discussions emphasized that AI4Science is not merely an incremental improvement over traditional simulation, but a shift in how scientific questions are formulated and explored. Key issues include the central role of data, computing power, and foundation models, the need for robust validation and trust in AI-generated knowledge, and the risks associated with the concentration of computational resources. At the same time, AI is increasingly viewed as a new scientific instrument, raising fundamental questions about reproducibility, interpretability, and the nature of discovery itself. In this rapidly evolving landscape, SCAI aims to position itself as a central hub for AI4Science at Sorbonne Université, fostering cross-disciplinary synergies and building strong links with national and international partners, as well as industry. The initiative seeks to accelerate AI-assisted discovery, scale up scientific AI models, and train a new generation of researchers and engineers capable of navigating both worlds: true “centaur scientists”. More broadly, it contributes to ensuring that France and Europe remain at the forefront of this transformation. The launch event made one point clear: AI is no longer just a tool for science. It is becoming an integral part of the scientific process itself. The challenge now is to structure this transformation scientifically, institutionally, and societally, to fully realize its full potential.
9:30 - Introduction (P. Cinnella, D'Alembert-SCAI) SLIDES 9:45 - Keynote 1 – Karthik Duraisamy (University of Michigan) "AI-Augmented Discovery Engines: Progress, Opportunities and Emerging Ecosystems" Scientific progress is entering a new phase: one where frontier reasoning models, domain foundation models, classical computational science, scientific instruments and human experts function as a tightly coupled agentic system. The real bar is not benchmark performance, but scientific utility: accelerating validated insight, compressing discovery and design cycles, and narrowing the sim-to-real gap in complex multiscale physics. I will describe recent advancements at the University of Michigan, highlighting (i) expressive, domain-specific foundation models for scientific prediction and inference, and (ii) agentic AI infrastructures that orchestrate tools, data, and HPC workflows for hypothesis generation, verification/ validation, and design optimization, while keeping humans in the loop to set objectives, constraints, and scientific judgment. I will also discuss the ambitious University of Michigan–Los Alamos National Laboratory superpartnership and place these efforts in the broader U.S. AI4Sciencelandscape, spanning public-private partnerships, national-scale infrastructure, and community-building initiatives. I will close by arguing that the next leap will come less from any single breakthrough model and more from the co-design of models, methods, instruments, and institutions. 10:30 - AI4Science at SCAI/Sorbonne Université - AI4Chemistry by Jean-Philip Piquemal (LCT) "A Quantum Foundation Machine Learning Model for Accurate Atomistic Simulations in Drug Design". While artificial intelligence has revolutionized the prediction of static protein structures, characterizing their dynamics and interactions with drug candidates remains a computational bottleneck. Here, the presentation will introduce FeNNix-Bio1 , a foundation machine learning model designed to power accurate, reactive atomistic simulations of biological systems at an unprecedented speed and scalability. Trained exclusively on synthetic quantum chemistry data, it accurately captures complex condensed-phase phenomena such as ion solvation and subtle liquid water properties for which it outperforms state-of-the-art specialized force fields. The presentation will also highlight the links between such methodology an other foundation models found in Biology, Chemistry and Material Science. - AI4Materials by Marco Saitta (LPENS) "Materials for Energy from AI-driven approaches at ENS and in France". SLIDES This presentation highlights recent AI-driven approaches to atomistic simulations in computational materials science and their impact on the study of energy materials, including batteries and supercapacitors. Selected results from our research group at ENS are discussed within the broader context of the French AI–Materials–Energy research ecosystem. - AI4Biology by Alessandra Carbone (LCQB) "Decoding protein function at scale: AI from structures to systems". SLIDES Deep learning has profoundly transformed biology by revealing the structure of the protein universe, with AlphaFold and related models providing atomic-level reconstructions at unprecedented scale. Yet structure is only the beginning. The next grand challenge is functional understanding: determining what proteins do, how they interact, and how they collectively give rise to biological processes.
11:30 Round table 1: AI4Science as a pathway to innovation. Panelists:
12:30–13:45 Buffet lunch 13:45 - Keynote 2 – Jesse Thaler (MIT, IHES): " Centaur Science: Adventures in AI+Physics". SLIDES The mythical centaur (half human, half horse) has become a metaphor for human-AI collaboration. In this talk, I explore what centaur science looks like at the intersection of artificial intelligence and fundamental physics. I share adventures from both directions of this exchange: teaching machines to "think like a physicist" by incorporating physics principles into machine learning frameworks, and teaching physicists to "think like a machine" to maximize discovery opportunities in both experimental and theoretical physics. 14:30 - AI4Science At SCAI/Sorbonne Université - AI4Climate by Claire Monteleoni (INRIA-ARCHES) "Generative AI for Climate Change". Many applications aimed at addressing climate change hinge on fundamental challenges of data fusion, interpolation, downscaling, and probabilistic domain alignment. I will provide a survey of our recent work developing generative AI methods for these problems, with applications including weather forecasting, climate model emulation and scenario interpolation, and renewable energy planning - AI4Astrophysics by Guilhem Lavaux (IAP) "Cosmic scales: integrating Bayesian inference and deep learning to reveal new physics". SLIDES Astronomy confronts the paradox of studying objects we can never touch, while ever‑improving, high‑fidelity data strain our simulation and inference pipelines. In this talk I will showcase several AI breakthroughs—Bayesian‑deep hybrids, neural surrogates, and probabilistic models—and the fresh discoveries they have enabled in exoplanet characterization, galaxy‑scale physics, and cosmological modeling across multiple observational regimes and theoretical frameworks - Physics-aware AI by Patrick Gallinari (ISIR) "AI for Physical Dynamics: From Neural Surrogates to Foundation Models". SLIDES 15:30 - Table ronde 2: AI for Science, Science for AI Panelists:
16:30 - Summary, recommendations, next steps and closing. 17:00 End of program (approx.)
Scientific committee:
|
