Presentation 1-1, 13:10~13:40.
Speaker
Hyerim Lee (Seoul National University)
Title
Inverse Computer Model Calibration Approach with Conditional Variational Auto-encoder
Abstract
Computer model calibration is essential for aligning simulation outputs with observational data, but traditional Bayesian approaches relying on Markov Chain Monte Carlo (MCMC) are often computationally prohibitive and difficult to scale for high-dimensional, dependent outputs. To address these challenges, we propose an inverse model-based calibration framework using Conditional Variational Autoencoders (CVAE) with novel latent structures, including stick-breaking and Dirichlet priors. Our approach enables efficient parameter estimation by learning inverse mappings from observations to input parameters while providing principled uncertainty quantification. To further improve coverage reliability, we integrate conformal prediction with studentized residual scores, correcting both undercoverage and overcoverage in posterior intervals. We demonstrate the framework on two case studies: calibrating the PSU3D-ICE model of the West Antarctic Ice Sheet using high-resolution Bedmap3 thickness data, and calibrating an elastic-decohesive sea ice model using satellite-derived crack data. Across both applications, CVAE-based models outperform baseline and MC-Dropout methods in balancing point estimation accuracy and uncertainty calibration, with stick-breaking and inverse-CDF variants showing particularly strong performance. These results highlight the potential of CVAE-based inverse calibration to handle complex spatial data and provide reliable uncertainty quantification, offering a scalable alternative to traditional Bayesian calibration.