dualis for Python

Coupled physics whose conservation audit is an exception you can catch.

pip install dualis        # wheels for linux, macos and windows; abi3, so 3.10 upward

import dualis

sim = dualis.Simulation(schedule="multirate", conservation_tolerance=1e-9)
sim.add_heater("element", watts=2.0, reserve_j=6.0)
sim.add_bar("bar", length_m=0.020, cells=41, area_m2=1e-4)

for _ in range(8):
    sim.advance(0.5)          # raises dualis.Violation if the books do not close

print(sim.temperature("bar"))     # 294.39 K
print(sim.profile("bar")[:3])     # warmer at the fed end

Several bodies and the drop between them

add_lump gives you one temperature for a whole assembly. The number that decides whether a motor survives is the winding, which is hotter than the case by however much the joint between them resists — so for that, a network:

sim = dualis.Simulation(schedule="staggered")
sim.add_heater("losses", watts=6.0, reserve_j=5_400.0)
sim.add_network("motor",
    nodes=[{"name": "winding", "material": "copper", "volume_m3": 18e-6,
            "thickness_m": 2e-3, "initial_k": 298.15},
           {"name": "case", "material": "aluminium", "volume_m3": 220e-6,
            "thickness_m": 4e-3, "initial_k": 298.15,
            "ambient_k": 298.15, "area_m2": 0.042}],   # both keys, or neither
    links=[{"from": "winding", "to": "case", "w_per_k": 0.9}],
    absorbing="winding")

for _ in range(900):
    sim.advance(1.0)

dict(sim.node_temperatures("motor"))     # {'winding': 311.2, 'case': 304.9}
sim.heat_flow_w("motor", "winding", "case")   # 5.585 W crossing the joint

material is any name in the kernel's catalogue, and this line no longer lists them — it used to say five, the binding accepted those same five against a catalogue of nine, and listing them here made a fourth copy of a spelling that had already gone stale three times. The error message names them, and it is generated from the catalogue, so it is right by construction. A node with ambient_k and area_m2 loses heat to still air; a node with neither is interior. Giving one without the other is refused rather than quietly producing a node that looks like it is cooling and is not.

Names are resolved once, when the network is built, so a link to a node that does not exist raises before any stepping happens and lists the nodes that do. That matters more here than it looks: a link adds +q to one node and −q to another in the same sum, so a missing or misdirected link passes the conservation audit at machine precision and simply reports the wrong temperature forever.

temperature() refuses a network rather than averaging it — the reason to build one is that its nodes differ, so a mean would describe no part of it.

Why this rather than numpy

Write physics in numpy and a wrong model runs happily, producing plausible output. There is no signal; you find out when somebody notices the answer is silly.

Every advance here either moves the clock or raises, and the exception is addressable rather than a sentence to parse:

try:
    sim.advance(0.5)
except dualis.Violation as v:
    v.quantity     # 'energy'
    v.site         # 'bus (published but not consumed)'
    v.before, v.after, v.scale, v.tolerance
    sim.time       # unchanged — a refused step does not advance the clock

Be clear about what it catches. Quantities appearing or vanishing, amounts left unclaimed on the bus, and a domain refusing a step it cannot take. It does not catch a model that is internally consistent and physically wrong. For that, check against something the library did not compute — which is what heat_capacity_j_per_k is exported for, and what tests/test_dualis.py does throughout.

What crosses the boundary, and what does not

Values cross as SI floats with the unit in the parameter name: length_m, watts, initial_k. dualis's dimensional types are a compile-time thing — Length + Time does not compile in Rust — and Python cannot have that. A runtime Quantity wrapper would cost an allocation and a check per operation to catch, at run time, a class of error a Python caller mostly does not make, because it passes values in and reads them out rather than doing algebra on them. So the types stay on the Rust side and the units stay in the names.

What crosses instead is the audit, which is the part worth having.

What this cannot do

You cannot run an ensemble from Python, and it is not an oversight. Ensemble exists to run many independent samples across threads, and a Python sample function cannot: the GIL serialises it. A binding for it would look parallel, measure slower than a plain Python loop because of the crossing cost, and have nothing left of the one thing it is for. Write the sampler in Rust, or loop in Python and accept one core.

You cannot write a domain in Python. That means calling back into the interpreter from inside the step loop, holding the GIL across it, and deciding what an exception raised mid-sweep does to a half-advanced simulation. All three are answerable and none cheaply, so this binds the physics the library has and says so rather than shipping a subtly wrong version of that.

So: enough to run and audit coupled physics from Python, not enough to extend it. The extending lives in Rust, and the workspace is where.

A Simulation is also bound to the thread that created it. Domain is not Send, and the binding says so rather than asking the kernel for a bound it does not offer.

There is no scene or view binding. The Rust side has dualis-scene for capturing a run and dualis-view for drawing one; neither crosses. Nothing forbids it in principle — the reason is that a Python caller already has matplotlib, pandas and every plotting stack there is, and what they want from this side is the numbers: profile, node_temperatures, temperature, ledger. Handing them an SVG instead would be the worse half of both worlds. If you want the HTML report, run the scene from Rust.

Building

Its own cargo workspace, deliberately. This one resolves fifteen external crates where the library resolves twelve, seven of them pyo3's and appearing nowhere in the library, plus a libpython link — and the library's lockfile, licence allow-list and WebAssembly jobs are promises that should not have to accommodate a Python extension.

cd bindings/python
pip install "maturin>=1.7,<2.0"
maturin build --release          # an abi3 wheel: one build serves 3.10 upward
pip install target/wheels/dualis-*.whl
python tests/test_dualis.py

maturin develop works too, in a virtualenv. CI builds the wheel, installs it and runs those tests on every commit — because "it compiles" proves neither half of "pip install then import".