Margam reference catalogue¶
For task-based help with current screenshots, start with the illustrated user manual and Your first PINN. F1 opens the built manual in the app.
Current implementation and recorded evidence, reviewed 28 October 2026. Start with setup and commands; historical design discussions are labelled below.
Using Margam¶
| Question | Read |
|---|---|
| Can I follow a screenshot walkthrough? | Your first PINN |
| How do I build, train, inspect and save a model? | Desktop guide |
| Which complete problems are available? | Examples and predefined PDEs |
| How do compact/expanded PINNs work? | Recipes |
| Which pointwise networks are available? | Neural networks |
| What does Margam mean by Expert PINN? | Expert’s Guide pipeline |
| How does learning-rate annealing work? | LRA method and safeguards |
| How do I integrate squared PDE/BC errors over the geometry? | Integral loss formulation |
| How do residual refinement and importance sampling work? | Adaptive sampling |
| How does SA-PINN work, and which settings matter? | SA-PINN method |
| Where is the original time-dependent Burgers case? | Burgers — 1D + time |
| What is the paper Helmholtz example? | Fourier/LRA Helmholtz |
| What is the steady heat problem/reference? | Heat demo |
| How do I export predictions to ParaView? | Explicit export and playback |
| How do prepared GeoSampler points enter Margam? | Planar data contract |
| What is implemented and what are the limits? | Component capability index |
Results and verification¶
| Question | Read |
|---|---|
| What numbers did we actually measure? | Results catalogue |
| Where are the exact records? | Curated evidence files |
| What do tests establish? | Validation scope and commands |
| How did hard Dirichlet compare with soft BCs? | Heat accuracy study |
A measured error, a test's acceptance threshold, and an implementation claim are different kinds of evidence. Result pages state the problem, method, seed, training budget, evaluation/reference, runtime and limitations. Historical GPU data does not replace testing the current CPU build on CUDA.
Extending the project¶
- Development and experiment commands
- Numerical API and benchmark runner
- Adding a PINN method and its one-file recipe
- Agent/repository instructions
- Remaining work
Keep these terms distinct:
| Term | Responsibility | Location |
|---|---|---|
| Method | Distinctive numerical behaviour/training | margam/methods/ |
| Recipe | Components, connections, defaults and exposed settings for one method construction | margam/recipes/ |
| Component | Reusable sampler, network, loss, operator or optimiser helper | Shared source modules; see capability index |
| Example | A physical problem wired to a method and explicit outputs/reference | margam/templates/; File → Examples |
Design history and provenance¶
Community methods and attribution credits DeepXDE for our RAR reference and NVIDIA Modulus for PINN importance sampling/integral loss.
Roadmap preserves the original discussions; its stage-specific “not implemented” notes are historical. Simulink evidence distinguishes sourced interactions from Margam choices. Cleanup record and GeoSampler reuse review are dated records, not the current capability catalogue. Third-party attribution remains separate.
When adding or changing a method, update its guide, capability row, example, validation evidence and results index together. Do not mark a proposed feature as implemented or a regression threshold as a measured benchmark result.
-
Evaluate supplied points: separate coordinate/reference blocks, preserved IDs, pointwise errors and VTU/PVD export.
-
Prepared training data: separate interior/boundary sources, free planar labels and a triangular mixed-BC example.
-
gPINN: scalar residual-gradient enhancement, coordinate weights, RAR and implementation evidence.
-
RBA-PINN: fixed-point residual-history attention, initialization modes and raw/weighted diagnostics.