Description
Propose adding a Neural Operator Factory (NOF) — a config-driven framework
for training neural operator surrogates for reservoir simulation and beyond. The NOF
supports 165 model architectures (FNO, U-FNO, Conv-FNO, FNO4D, DeepONet
with 8 variants including TNO), autoregressive training with three-stage
pipelines, and physics-informed losses, all configurable from YAML.
Motivation
Reservoir simulation is a key application for neural operators, but
experimenting across architectures (FNO vs DeepONet vs TNO) currently
requires separate codebases. The NOF unifies these under a single
training pipeline with shared data loading, loss functions, masking,
and distributed training infrastructure.
Scope
- New example under
examples/reservoir_simulation/neural_operator_factory/
- No changes to PhysicsNeMo core library
- Minor additions: S101 per-file-ignore in pyproject.toml,
interrogate baseline update, CHANGELOG entry
- Includes 375 unit tests and 5 reproducible examples from published papers.
Description
Propose adding a Neural Operator Factory (NOF) — a config-driven framework
for training neural operator surrogates for reservoir simulation and beyond. The NOF
supports 165 model architectures (FNO, U-FNO, Conv-FNO, FNO4D, DeepONet
with 8 variants including TNO), autoregressive training with three-stage
pipelines, and physics-informed losses, all configurable from YAML.
Motivation
Reservoir simulation is a key application for neural operators, but
experimenting across architectures (FNO vs DeepONet vs TNO) currently
requires separate codebases. The NOF unifies these under a single
training pipeline with shared data loading, loss functions, masking,
and distributed training infrastructure.
Scope
examples/reservoir_simulation/neural_operator_factory/interrogate baseline update, CHANGELOG entry