▣ Title : Towards foundation models for discovery and optimization in nuclear energy systems

▣ Speaker : Prof. Ricardo Vinuesa(University of Michigan)

▣ Date & Time : 2026. 4. 23(Thu) 16:30~17:30

▣ Venue : Research Bldg I, #310

▣ Language : English

▣ Abstract :

Towards foundation models for discovery and optimization in nuclear energy systems Ricardo Vinuesa(University  of  Michigan),  Elia Merzari(Penn State) and Paul  Fischer(University of Illinois, Urbana) Next-generation nuclear energy systems(spanning advanced fission reactors and magnetic-confinement fusion devices) depend on our ability to understand, predict and control tightly coupled  multiphysics phenomena.  These  include  turbulent  thermos-hydraulics, neutronics-thermal feedback, material degradation under extreme conditions, and plasma-wall interactions. Each operates across a wide range of spatial and temporal scales and is governed by nonlinear partial differential equations. Despite decades of progress, the design and  safety  analyses  of  nuclear  systems  still  rely  on  computationally expensive, case-specific high-fidelity simulations and experiments that produce siloed datasets with limited transferability  across  geometries, regimes,  and  reactor  concepts. This  fragmentation significantly  slows  discovery,  limits  uncertainty-aware  optimization,  and  constrains innovation in both fission and fusion technologies. 

  We propose a transformative foundation-model framework for nuclear energy that learns universal, causal representations of multiphysics behavior across reactor and confinement concepts. The core innovation is a geometry-agnostic formulation in which all simulation and experimental data (regardless of mesh topology, reactor layout or operating regime) are mapped onto a common  reference  domain.  Thermo-hydraulic fields,  neutron fiuxes, material state variables or plasma quantities are encoded together with their governing parameters (e.g., Reynolds  and  Prandtl  numbers,  power  density,  material  properties, magnetic field strength) into a unified latent space. In this representation, the learns how geometry and operating conditions modulate nuclear-relevant physics in a generalizable manner.

  The  architecture  integrates  three  complementary  generative  components.  First, a β-variational  autoencoder  compresses  high-dimensional  multiphysics  snapshots  into a compact latent representation capturing dominant physical structures. Second, a temporal transformer models the evolution of these latent states, enabling the capture of transients, instabilities, and long-range temporal dependencies relevant to reactor safety and plasma dynamics.  Third, conditional  latent diffusion  learns  the  probability  distribution across operating regimes and geometries, allowing the model to synthesize physically consistent states for previously untested reactor configurations or plasma conditions. This capability enables rapid exploration of design and safety envelopes at orders-of-magnitude lower cost than traditional HPC workflows.

  Crucially, the framework is designed for scientific discovery. Causal-inference tools are applied directly within the latent space to identify regime transitions, triggering mechanisms for instabilities, and key feedback pathways(e.g., thermal-hydraulic-neutronic coupling or plasma confinement degradation). These causal insights are then mapped back to physical space using explainable AI, revealing the spatial regions and temporal events responsible for the identified behavior. The result is a model that produces interpretable, mechanistic understanding alongside fast predictions.

  When  coupled with an agentic AI system, the foundation  model  enables  autonomous exploration of the nuclear design space. Operating entirely in the low-dimensional latent manifold, the agent can search for configurations that optimize safety margins, efficiency, or confinement performance, while uncertainty quantification guides where additional high-fidelity simulations or experiments are most valuable. This closed-loop workflow continuously improves model fidelity while accelerating discovery.

By unifying heterogeneous nuclear multiphysics into a single, interpretable, and actionable model, this approach promises to dramatically shorten design cycles, reveal previously inaccessible  scaling  laws,  and  enhance  the  safety  and  performance  of  both  advanced fission and fusion systems.