Dynamics only

A 512-km basin has land on its last x and y grid lines, rotating ocean currents and the reference’s fixed wind snapshot 15 (day 5.875). Wind stays constant throughout the run. Defaults are 1024 by 1024 physical cells, 100 steps of 600 seconds and 120 adaptive EVP iterations per step. EVP, or elastic-viscous-plastic, is the method used to solve for ice velocity and stress (alpha/beta 500, absolute wind forcing, which uses wind velocity without subtracting ice velocity). This case has no thermodynamic growth, meaning freezing or melting. A small local CPU run is:

python -m veris.setups.run_dyn --nx 32 --ny 48 --steps 2 --evp-steps 4 \
    --backend cpu --output-snapshot output/dynamics.nc

Use --backend gpu for a GPU directly on the current node. If the environment already sets JAX_PLATFORMS=cpu, unset it or set JAX_PLATFORMS=cuda first. The output contains all final State fields with halos removed, in (x, y) order. State is the collection of model arrays. Halos are extra cells that hold boundary or neighboring data. The runner reports compilation and integration times separately. Compilation prepares executable code for the required loop shapes without a warmup model run. Integration executes exactly the requested model steps.

The command-line runner uses Experiment.builtin("dynamics", ...) and Simulation for the same run process as coupled (motion plus growth) and growth experiments. The definition holds global physical dimensions and settings. Forcing (prescribed inputs), time and output use the shared Simulation lifecycle interface. A public Python experiment is:

from veris import Execution, Experiment, Grid, Simulation

experiment = Experiment.builtin(
    "dynamics", grid=Grid(32, 48), steps=3,
    settings={"dtype": "float32", "nEVPsteps": 4},
    execution=Execution(backend="cpu"),
)
with Simulation.initialize(experiment) as simulation:
    result = simulation.run()

The shared initializer returns initial fields and model objects that stay fixed during compilation:

import jax
from functools import partial
from veris import Experiment, Grid, advance, initialize_experiment, step

jax.config.update("jax_enable_x64", True)
state, settings, constants = initialize_experiment(
    Experiment.builtin("dynamics", grid=Grid(32, 48))
)
transition = partial(
    advance, configuration=settings, constants=constants, physics="dynamics",
)
state = step(state, transition, 3, has_aux=True)

advance returns State and Diagnostics. The public veris.step repeats timesteps with checkpointing, which recalculates intermediate values to save memory during reverse differentiation. See Scan rollouts and checkpointing for forcing, diagnostic histories and runs split across devices. Ocean stresses are separate Diagnostics fields. Outputs for exchange with an ocean model do not add fields to State. Experiment settings, constants and scenario groups let you control experiments without changing source files.

Reference adaptations

The runnable cases correct the old serial mask’s five-cell neighbor offset. Masks identify ocean and land. The cases use the physical one-cell face stencil, a calculation based on neighboring cells, and initialize periodic halos consistently. They replace generated initializer imports, fixed home paths and file writes by every process. Prescribed fields keep their reference values. Fields not prescribed, including ocean and air temperature, start at zero as in initialize_dyn.py. Dynamics-only execution does not use those temperatures. Independent data from the original source checks every startup field, shared numerical setting and physical constant. It also checks physical fields after 100 timesteps with 120 adaptive EVP iterations on a reduced 16 by 16 grid. Experiment-specific constants live in scenario registries; universal physical constants stay in PhysicalConstants. This requires no changes to the physics calculations.

Derivative validation

The exact dynamics initial state has zero strain (no ice deformation) and closed walls. Safeguarded magnitude calculations give a finite chosen local linear response there. Tests check the actual initial State, its full reverse derivative and changing wind sensitivities. Finite differences, which compare results after small input changes, check independent smooth reference calculations and the growth column’s sensitivity to longwave radiation. See Automatic differentiation at model boundaries for choices at zero magnitude, physical branch thresholds and the special free-drift case with an unbounded sensitivity.

Scenario controls

Dynamics only scenario controls

Setting	Default	Units	Description
length	512000.0	m	Basin length in each horizontal direction
wind_time	5.875	day	Fixed reference wind snapshot time
wind_speed	15.0	m s-1	Wind amplitude
wind_angle	1.2566370614359172	rad	Wind rotation angle
wind_scale	50000.0	m	Wind radial normalization length
wind_decay	100000.0	m	Wind exponential decay length
center_offset	0.1	1	Initial wind center as basin fraction
center_rate	0.1	day-1	Wind center translation as basin fraction per day
ocean_speed	0.01	m s-1	Ocean circulation speed scale
depth	1000.0	m	Ocean basin depth
ice_thickness	0.3	m	Initial ice thickness including wall cells
surface_temperature	273.0	K	Initial ice surface temperature
coriolis_start	0.00014604	s-1	Coriolis frequency at first y cell
coriolis_offset	0.00014596	s-1	Reference last-cell Coriolis intercept
coriolis_increment	8e-08	s-1	Reference Coriolis increment per y cell