Beyond hierarchies: generalized structural exploitation in multi-fidelity sampling and learning
Alex Gorodetsky
University of Michigan
Aerospace Engineering
Computational science is increasingly driven by the need to fuse information from diverse sources—ranging from high-fidelity simulations and experimental data to low-order approximations and data-driven models. While multi-fidelity methods have become a standard tool for reducing the cost of many-query problems, their success relies heavily on how effectively they exploit the underlying mathematical structure connecting these information sources.
In this talk, I will categorize and explore the dominant structural mechanisms that enable efficient information fusion. We will discuss how sampling methods exploit statistical correlations for variance reduction, how surrogate strategies leverage residual learning to correct low-fidelity baselines, and how solver-intrusive approaches utilize temporal and spatial error structures directly within the simulation loop. I will then focus on how these structures can be encoded via graphical relationships. Here we will focus on our recent work with MFNets (Multi-Fidelity Networks), a framework that generalizes these relationships beyond rigid hierarchies. By treating model fusion as a problem of inference on Directed Acyclic Graphs (DAGs), MFNets allow for "all-at-once" learning of complex, non-nested systems where information sources may share partial inputs or peer-to-peer dependencies.
Finally, I will present work-in-progress applying these architectures to complex Laser-Plasma Interaction (LPI) simulations. LPI represents a challenge for multi-fidelity methods due to the chaotic nature of the physics and the difficulty in establishing stable correlations across fidelities. I will discuss our current approaches for surrogate-based fusion in this domain, highlighting both preliminary results and the open structural challenges.