Releases: phydra-labs/phydrax
Phydrax v0.2.2
What's Changed
- feat(nn): add Sequential model wrapper by @leonard-gleyzer in #15
- v0.2.2 by @leonard-gleyzer in #16
Full Changelog: v0.2.1...v0.2.2
Phydrax v0.2.1
What's Changed
- Add scan option across models and introduce SeparableKAN by @leonard-gleyzer in #10
- Add wrapper scan paths and differential loop optimizations by @leonard-gleyzer in #11
- Fix Windows temp-folder handling for geometry CAD intermediates by @leonard-gleyzer in #12
- feat(constraints): support DomainFunction weights for residual constraints by @leonard-gleyzer in #13
- v0.2.1 by @leonard-gleyzer in #14
Full Changelog: v0.2.0...v0.2.1
v0.2.0
Phydrax v0.2.0
Phydrax 0.2.0 is a major ergonomics and performance release across sampling, differentiation, PCI overlays, and examples.
Highlights
1) Unified sampling API (num_points)
Constraint sampling now uses a single interface:
- Dense: num_points=1024 or num_points=(256, 256)
- Coord-separable: num_points={"x": 128, "t": 256}
- Mixed dense + separable: num_points=(8, {"x": axis_spec})
This removes split configuration and simplifies sampling setup.
2) Differential operator execution upgrades
- Explicit AD engine behavior (including JVP path) where applicable
This improves consistency and helps optimized routes apply to structured/
factorized models.
Physics-Constrained Interpolation (PCI) enforced overlays now better
preserve optimized derivative paths through composition, including:
- Derivative-aware blend/quotient-style composition
- Improved boundary/initial compatibility handling
- Better propagation of optimized derivative behavior through enforced ansätze
4) New example: examples/wave1d.py (Marimo notebook)
Added a full tutorial wave equation example demonstrating:
- Latent-factorized modeling
- PCI-enforced BC/IC overlays
Public notebook:
5) Docs/site updates
-
Added Docs → Examples → Overview
-
Added LinkedIn docs branding/icon updates
-
Updated guides/API docs for new sampling and differentiation behavior
-
Version bump: 0.1.1 → 0.2.0
-
Notebook deps surfaced for example workflow (marimo, matplotlib)
-
Solver training logs include per-iteration timing
Migration notes
Sampling API
If you previously used separate coord-separable sampling arguments, migrate to num_points forms.
Before:
num_points=8
coord_separable={"x": phx.domain.UniformAxisSpec(nx)}
dense_structure=phx.domain.ProductStructure((("data",),))After:
num_points=(8, {"x": phx.domain.UniformAxisSpec(nx)})
dense_structure=phx.domain.ProductStructure((("data",),))Or fully separable:
num_points={"x": 128, "t": 256}Validation scope
This release includes broad test updates/additions across:
- Constraint sampling behavior and edge cases
- Differential operators/runtime cache behavior
- Latent contraction structured/flat behavior
- Enforced solver/overlay behavior
- Docs example execution path
Full Changelog: v0.1.1...v0.2.0
Initial release!
Initial release