MultiFidelity

Multifidelity is an approach in computational science and engineering that combines information from models or data sources with different levels of accuracy and cost. The central idea is to use inexpensive, low-fidelity models to guide and accelerate the use of expensive, high-fidelity models, achieving accurate predictions with reduced computational expense.

Common techniques include multifidelity surrogate modeling, such as co-kriging or autoregressive models that relate high-fidelity outputs to low-fidelity ones plus a correction term. Kennedy and O’Hagan introduced the co-kriging framework, while multifidelity Monte Carlo methods use control variates across fidelities to reduce variance. Other methods include MFMC, multi-fidelity optimization, and Bayesian calibration.

Applications span aerospace, automotive, energy, and physical sciences. They are used in design optimization, uncertainty quantification, parameter estimation, and machine learning where simulations or experiments are costly. In practice, a workflow may build a hierarchy of models, learn a statistical relationship between fidelities, and progressively sample the most informative points across levels.

Benefits include substantial cost savings and faster convergence, as long as fidelities are correlated and the discrepancy between levels can be modeled. Limitations include model misspecification, diminishing returns with poorly correlated models, and added methodological complexity or overhead.