Multifidelity Model Calibration in Structural Dynamics Using Stochastic Variational Inference on Manifolds.

Area of Science:

• Engineering
• Computational Statistics
• Machine Learning

Background:

• Bayesian techniques using Gaussian processes (GPs) excel at uncertainty quantification and data efficiency in engineering.
• However, traditional GP methods face computational challenges with large datasets and numerous inputs.
• This limits their practical application in complex engineering scenarios.

Purpose of the Study:

• To enhance Gaussian process (GP)-based metamodeling and model calibration for large-scale engineering problems.
• To address the computational intractability of standard GP methods with extensive training data.
• To enable efficient and rigorous Bayesian inference in data-rich environments.

Main Methods:

• Employed a stochastic variational inference algorithm for Gaussian process (GP) metamodeling.
• Applied the algorithm to accelerate statistical learning of calibration parameters and hyperparameter tuning.
• Focused on retaining the precision of Bayesian inference despite computational demands.

Main Results:

• Demonstrated significant improvements in computational performance for GP-based tasks.
• Successfully applied the algorithm to metamodeling and model calibration with thousands of data points.
• Validated the method's effectiveness on multiple complex engineering problems.

Conclusions:

• The stochastic variational inference algorithm offers a computationally efficient solution for GP metamodeling and calibration.
• This approach makes advanced Bayesian techniques more feasible for large-scale engineering applications.
• The method successfully balances computational speed with the rigor of Bayesian inference.