[BUG] - Bug in Example 9-Unsteady_Flow: Swarm.advection(order=2) fails after dynamic particle addition

[YeYangLiu]

Describe the bug

While running the official example 9-Unsteady_Flow.ipynb, I encountered a ValueError during the advection step. The error occurs when using order=2 (RK2) for swarm advection after adding new particles to the swarm. It seems the internal buffers for the RK2 scheme do not resize to accommodate the updated swarm size.

To Reproduce

Python
import underworld3 as uw

passive_swarm = uw.swarm.Swarm(mesh=mesh)
passive_swarm.populate(
    fill_param=0,
)
max_steps = 50
timestep = 0
elapsed_time = 0.0
delta_t = 0.05

for step in range(0, max_steps):

    # Keep previous guess for solve
    navier_stokes.solve(zero_init_guess=False, timestep=delta_t)
    passive_swarm.advection(v_soln.sym, delta_t, order=2, corrector=False)

    # Add new tracer particles near the inlet
    new_points = 100
    new_coords = np.array([0.0, 0.25] + 0.5 * np.random.random((new_points, 2)))
    passive_swarm.add_particles_with_coordinates(new_coords)
    ...

The following snippet is a simplified version of the logic in 9-Unsteady_Flow:

## --- Step 1: Initial advection (works fine) ---
# This likely initializes internal buffers with the current swarm size
passive_swarm.advection(v_soln.sym, delta_t, order=2)

## --- Step 2: Add new particles (as seen in Example 9) ---
new_points = 100
new_coords = np.array([0.0, 0.25] + 0.5 * np.random.random((new_points, 2)))
passive_swarm.add_particles_with_coordinates(new_coords)

## --- Step 3: Second advection (Crashes) ---
# This fails with a shape mismatch error
passive_swarm.advection(v_soln.sym, delta_t, order=2)

## Error Message
ValueError: could not broadcast input array from shape (N_old, 2) into shape (N_new, 2)
(Note: In my case, it was (742, 2) and (842, 2))

Environment:

OS: macOS (Apple Silicon M3)

Underworld3 Version: [0.99.0b]

Example: 9-Unsteady_Flow.ipynb

[lmoresi]

@YeYangLiu — thanks for the clear reproduction, this was straightforward to track down.

Root cause: add_particles_with_coordinates() resizes the underlying PETSc swarm fields but didn't invalidate the cached numpy arrays used by SwarmVariables. The RK2 advection buffer (_X0) retained its old-size array, causing the shape mismatch on the next advection(order=2) call.

Fix: commit 07d61f7 on the development branch adds the missing cache invalidation.

How to test

If you're using pixi (recommended):

git clone https://github.com/underworldcode/underworld3.git
cd underworld3
git checkout development
pixi run build
pixi run python your_script.py

Or if you already have the repo cloned:

cd underworld3
git fetch origin
git checkout development
git pull
pixi run build
pixi run python your_script.py

Could you confirm the fix resolves the issue in your setup? In particular, the 9-Unsteady_Flow.ipynb workflow with order=2 advection after add_particles_with_coordinates.

[YeYangLiu]

Hi @lmoresi,

I’ve confirmed that the fix resolves the issue!

Just a small note on the workflow: in my setup, simply running pixi run build after applying the changes didn't seem to sync the updates to the virtual environment. To resolve this, I removed the .pixi directory and reconfigured the environment using ./uw setup. After this fresh setup, the bug was successfully fixed.

The 9-Unsteady_Flow.ipynb workflow is now running perfectly with order=2 advection and dynamic particle addition on my Mac M3. Thank you again for the quick and effective support!