physure 0.2.0

High-performance physical dimension engine. Rust-first, zero-copy FFI. Successor to physure.

physure 🔬

Unit-aware, dimension-correct computing for Python — powered by a Rust core

Units, dimensions, and correlated uncertainty tracked through every calculation, with zero overhead under torch.compile / jax.jit.

---

Why physure?

Most unit libraries make you choose between correctness and speed. physure (physics + measure) refuses the trade-off:

• Correlated uncertainty propagation. Full sparse-covariance tracking between quantities (GUM-style), not just independent error bars. If x and y share history, x - y knows it.
• Native-speed dimensional analysis. Unit arithmetic runs in Rust (~50 ns per operation) with rational exponents — no floating-point drift in dimensions, zero-copy buffer FFI to NumPy.
• ML-ready. Quantity wraps NumPy, PyTorch, and JAX arrays. Under @physure.jit, units are validated at trace time and evaporate at runtime: ~1.17× vs. raw compiled PyTorch.
• Static unit checking. A mypy plugin narrows Q_(3, "m/s") to Quantity[..., Literal["m/s"]], so unit mismatches can fail before your code even runs.
• Zero runtime dependencies. pip install physure pulls in nothing else. NumPy/PyTorch/JAX/pandas support activates automatically when those packages are present.

Quick start

from physure import Q_

d = Q_(10, "km")
t = Q_(2, "hr")
print((d / t).to("m/s"))    # 1.3888888888888888 m/s

# Uncertainty propagates automatically — correlations included
g = Q_(9.8, "m/s^2", uncertainty=0.02)
m = Q_(2.5, "kg", uncertainty=0.001)
E = m * g * Q_(12, "m")
print(E.to("J"))            # (294.0 ± 0.61) J

Or straight from the command line — physure ships a unit-aware calculator and REPL:

$ python -m physure "500 N / 2 m^2 => kPa"
0.25 kPa

Highlights

Units that vanish at compile time

@physure.jit traces your function once, validates every dimension in Rust, then runs on raw tensors — dimensional safety with no per-call cost:

from physure import Q_, jit

@jit
def kinetic_energy(mass, velocity):
    return 0.5 * mass * velocity**2

kinetic_energy(Q_(10.0, "kg"), Q_(5.0, "m/s"))   # 125.0 kg·m²/s²
kinetic_energy(Q_(1.0, "m"), Q_(1.0, "s"))       # raises at trace time: incompatible units

Works with plain floats, NumPy arrays, PyTorch tensors (via __torch_dispatch__ + torch.compile), and JAX (jax.jit).

Uncertainty done properly

Choose the propagation mode globally or per-block:

import physure

physure.propagation_mode("correlated")     # full covariance (default: uncorrelated)

with physure.uncertainty_mode("uncorrelated"):
    ...                                    # scoped override

Backends include Gaussian (first-order), Monte Carlo, and Unscented Transform. Covariance lives in a sparse Rust store, so large lineages stay fast and memory stays flat.

Batteries included, loaded lazily

Pandas ExtensionArray, pydantic validation, SymPy symbolic quantities, unit-aware torch.nn layers, Arrow IPC serialization, plotting helpers, and a physics-as-text DSL (stress = 500 N / 2 m^2) — each activates only when you use it. Cold import stays around 20 ms.

How it compares

	physure	pint	astropy.units	unyt
Correlated uncertainty (covariance)	✅	—	—	—
Built-in uncertainty propagation	✅	via uncertainties	limited	—
Rust-accelerated core	✅	—	—	—
torch.compile / jax.jit compatible	✅	—	—	—
Static unit checking (mypy)	✅	—	—	—
Runtime dependencies	none	none	astropy stack	numpy
Ecosystem maturity	new	✅ mature	✅ mature	mature

If you need a battle-tested converter with a decade of integrations, pint and astropy are excellent. physure is for when you also need uncertainty you can defend and units inside compiled ML code.

Performance

Benchmark	Result
Cold import	~21 ms
Unit multiply/divide (Rust core)	~54 ns
Scalar add with dimension check	~40 ns
10⁶-element tensor op, @torch.compile	1.17× vs. pure PyTorch
Covariance propagation (sparse blocks)	~7 µs

Full methodology and reproduction steps: BENCHMARKS.md.

Installation

pip install physure            # Rust-compiled wheel, no other dependencies
pip install "physure[numpy]"   # + NumPy/SciPy/Numba acceleration
pip install "physure[torch]"   # + PyTorch backend
pip install "physure[jax]"     # + JAX backend
pip install "physure[all]"     # everything

From source

git clone https://github.com/Alexisrx96/physure
cd physure
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh   # Rust, if needed
maturin develop --release
uv sync --group dev

Architecture

physure/                     # Cargo workspace root
├── physure-core/            # 🦀 Pure Rust physics engine — no FFI deps
│   └── src/                 #    units, quantity, covariance, uncertainty,
│                            #    symbolic, Arrow serialization
└── physure-python/          # 🐍 PyO3 bindings + Python application layer
    └── physure/
        ├── domain/          # Quantity, units, dimensions, uncertainty
        ├── application/     # Q_ factory, unit-system context, startup
        ├── backends/        # NumPy / PyTorch / JAX adapters
        ├── _jit/            # tracing + compile-time dimension checks
        ├── ext/             # grammar DSL, IO, pandas, numba
        └── nn/              # unit-aware neural network layers

The rule: physure-core is the single source of truth. All math lives in Rust; Python is a thin, zero-copy translation layer.

Documentation

• Unit reference — every unit, prefix, and constant
• Tutorials and examples — including a unit-checked PINN notebook
• torch.compile integration

Contributing

Issues and PRs are welcome. The quality bar is enforced in CI: ruff clean, tests green on Python 3.11–3.14 with ≥ 80 % coverage, and zero new SonarQube violations. See CLAUDE.md for the full development guide.

License

MIT — Irvin Torres