RegAE.jl
Inverse Modeling • 2020 • Active
RegAE.jl provides a VAE-based approach for regularizing inverse problems and reducing parameter-space complexity while preserving coherent structures.
RegAE.jl provides an efficient way to regularize inverse problems where parameter fields are high-dimensional and exhibit coherent structures.
The approach leverages a variational autoencoder to learn useful regularization behavior from unconditioned realizations without requiring prior knowledge of the governing physics.
By easing the regularization process and reducing the dimensionality of the optimized parameter space, RegAE.jl can make inverse analyses more computationally tractable.
Adjoint methods and automatic differentiation can be used to further reduce computational cost during inverse analysis workflows.
Capabilities
- VAE-based regularization for inverse problems
- Handles coherent high-dimensional parameter fields
- Supports lower-cost optimization through dimensionality reduction