Talk II -A tutorial on multi-fidelity methods for fusion energy research
Ionut-Gabriel Farcas
Virginia Tech
Informatics
Multi-fidelity methods leverage models and data sources of varying cost and accuracy to enable efficient computation in many-query settings. In this two-part tutorial, we introduce the fundamental ideas behind multi-fidelity modeling and discuss their emerging applications in fusion research. We begin with basic terminology and concepts, followed by an overview of multi-fidelity approaches for outer-loop applications, including uncertainty quantification, inverse problems, optimization, and control. We then focus on multi-fidelity Monte Carlo sampling, which employs hierarchies of high-fidelity and lower-cost models to reduce computational cost while maintaining predictive accuracy. Next, we discuss a context-aware extension of multi-fidelity Monte Carlo that optimally balances the construction of low-fidelity models with their use in sampling. In this setting, we show that low-fidelity models can in some cases be too accurate for optimal multi-fidelity sampling, highlighting a key distinction from traditional model reduction, where increasing reduced-model accuracy directly improves the final result. Throughout the tutorial, we illustrate these ideas with examples from plasma micro-instability simulations in tokamak configurations. We conclude with a broader discussion of multi-fidelity approaches for optimization, inverse problems, and approximation, and highlight several open challenges in applying these methods to fusion science.