Description
Propose adding xDeepONet — the extended DeepONet family — to
physicsnemo.experimental.models.xdeeponet. The package provides a
config-driven, unified implementation of eight DeepONet-based
operator-learning architectures for both 2D and 3D spatial domains:
deeponet, u_deeponet, fourier_deeponet, conv_deeponet,
hybrid_deeponet — single-branch variantsmionet, fourier_mionet — two-branch multi-input variantstno — Temporal Neural Operator (branch2 = previous solution)
Motivation
DeepONet is a foundational neural-operator architecture with no current
first-class implementation in PhysicsNeMo. This contribution upstreams
a production-tested, config-driven xDeepONet family from the Neural
Operator Factory (#1551)
into the library proper, making it available to all PhysicsNeMo users
and not just reservoir-simulation practitioners.
Scope
- New package:
physicsnemo.experimental.models.xdeeponet
- Composable Fourier/UNet/Conv spatial branches (reuses
physicsnemo.nn.SpectralConv2d/3d and physicsnemo.models.unet.UNet)
- Three decoder types:
mlp, conv, temporal_projection
- Automatic spatial padding, trunk coordinate extraction,
adaptive-resolution pooling
- 29 unit tests under
test/experimental/models/test_xdeeponet.py
- CHANGELOG entry
- Package README with variants table and usage examples
Discussion
Design agreed with @coreyjadams, @peterdsharpe, and @ram-cherukuri in
Slack on. Custom UNet dropped in favour of the library's
physicsnemo.models.unet.UNet; placed under experimental/ per the
new-model convention.
Related
Description
Propose adding xDeepONet — the extended DeepONet family — to
physicsnemo.experimental.models.xdeeponet. The package provides aconfig-driven, unified implementation of eight DeepONet-based
operator-learning architectures for both 2D and 3D spatial domains:
deeponet,u_deeponet,fourier_deeponet,conv_deeponet,hybrid_deeponet— single-branch variantsmionet,fourier_mionet— two-branch multi-input variantstno— Temporal Neural Operator (branch2 = previous solution)Motivation
DeepONet is a foundational neural-operator architecture with no current
first-class implementation in PhysicsNeMo. This contribution upstreams
a production-tested, config-driven xDeepONet family from the Neural
Operator Factory (#1551)
into the library proper, making it available to all PhysicsNeMo users
and not just reservoir-simulation practitioners.
Scope
physicsnemo.experimental.models.xdeeponetphysicsnemo.nn.SpectralConv2d/3dandphysicsnemo.models.unet.UNet)mlp,conv,temporal_projectionadaptive-resolution pooling
test/experimental/models/test_xdeeponet.pyDiscussion
Design agreed with @coreyjadams, @peterdsharpe, and @ram-cherukuri in
Slack on. Custom UNet dropped in favour of the library's
physicsnemo.models.unet.UNet; placed underexperimental/per thenew-model convention.
Related