Skip to content

Model Architectures

Deep-UQ separates predictive architectures from uncertainty methods. This page is the canonical inventory of model families available in the package, while the UQ method pages describe how posterior or predictive uncertainty is computed.

For convolutional backbones, the primary UQ paths are MC Dropout and Deep Ensembles. For coordinate-input PINNs, the primary UQ path is last-layer Laplace because the current Laplace implementation integrates naturally with final nn.Linear heads.

Legend: supported, not the primary path.

Dense / Parametric Models

These models operate on vectors or coordinate-augmented point inputs and are a natural fit for classical regression baselines, ensembles, and PINN-style training.

Model 1D 2D 3D Laplace MC Dropout Deep Ensemble Primary Use Notebook
MLP Parametric regression / classification Deep Ensembles
PINN1D 1D physics-informed PDE surrogate PINN + Poisson
PINN2D 2D physics-informed PDE surrogate PINN + Poisson

Spatial Convolutional Models

These are field-to-field backbones for regular scientific grids. Their main UQ story in Deep-UQ is stochastic inference through dropout or population-level uncertainty through ensembles.

Model 1D 2D 3D Laplace MC Dropout Deep Ensemble Primary Use Notebook
CNNRegressor2D Compact 2D PDE surrogate baseline CNN / ResNet + Heat2D
ResNetRegressor2D Stronger residual 2D surrogate baseline CNN / ResNet + Heat2D
UNet2D Multi-scale 2D field-to-field surrogate UNet + Diffusion2D
UNet3D Volumetric 3D field surrogate backbone UNet + Diffusion2D

Operator-Learning Models

These models learn mappings between whole fields. DeepONet uses separate branch and trunk networks, while FNO performs global spectral mixing on regular grids.

Model 1D 2D 3D Laplace MC Dropout Deep Ensemble Primary Use Notebook
DeepONet1D 1D field-to-field operator learning DeepONet + 1D Poisson
DeepONet2D 2D field-to-field operator learning DeepONet + Burgers
FNO2D 2D Darcy / regular-grid PDE surrogate FNO2D + Darcy
FNO3D 3D periodic PDE surrogate FNO3D + 3D Heat

Graph Operator Models

These models treat a regular grid as a graph and use local message passing instead of global spectral mixing. They are a good bridge between grid-based operator learning and future unstructured-mesh models.

Model 1D 2D 3D Laplace MC Dropout Deep Ensemble Primary Use Notebook
GraphNeuralOperator2D Grid-as-graph operator learning for reaction-diffusion and mesh-style surrogates Graph Operator + Gray-Scott

Generative / Diffusion Models

These models represent uncertainty through the spread of generated conditional samples rather than by wrapping a deterministic predictor with Laplace, dropout, or ensembles.

Model 1D 2D 3D Laplace MC Dropout Deep Ensemble Primary Use Notebook
ConditionalUNet2D Conditional field reconstruction with sample-based UQ Conditional Diffusion + Heat2D

Gaussian Process Models

The GP family is function-space Bayesian from the start, so the uncertainty story is native rather than layered on after training.

Model 1D 2D 3D Laplace MC Dropout Deep Ensemble Primary Use Notebook
GP regressors / classifiers Kernel-based Bayesian modeling Gaussian Processes

Deep Ensemble Method Note

DeepEnsembleWrapper is a UQ method, not a predictive model class. It can wrap any compatible deterministic backbone and is the primary ensemble-based uncertainty path for MLP, CNNRegressor2D, ResNetRegressor2D, UNet2D, and UNet3D.