JaxPM/notebooks/02-Advanced_usage.ipynb
Francois Lanusse 70ab9f1931 remove deprecated stuff
Former-commit-id: da6fdea0556407c968dd31d44af832c997f2645c
2024-12-21 09:58:42 -05:00

5.1 MiB
Raw Blame History

Advanced Particle Mesh Simulation on a Single GPU

Open In Colab

In [ ]:
!pip install --quiet git+https://github.com/DifferentiableUniverseInitiative/JaxPM.git
!pip install diffrax
In [ ]:
import jax
import jax.numpy as jnp
import jax_cosmo as jc

from jaxpm.painting import cic_paint , cic_paint_dx
from jaxpm.pm import linear_field, lpt, make_diffrax_ode
from jaxpm.distributed import uniform_particles
from diffrax import ConstantStepSize, LeapfrogMidpoint, ODETerm, SaveAt, diffeqsolve

Particle Mesh Simulation with Diffrax Leapfrog Solver

In this setup, we use the LeapfrogMidpoint solver from the diffrax library to evolve particle displacements over time in our Particle Mesh simulation. The novelty here is the use of a Leapfrog solver from diffrax for efficient, memory-saving time integration.

  • Leapfrog Integration: This symplectic integrator is well-suited for simulations of gravitational dynamics, preserving energy over long timescales and allowing larger time steps without sacrificing accuracy.
  • Efficient Displacement Tracking: We initialize only displacements (dx) rather than absolute positions, which, combined with Leapfrogs stability, enhances memory efficiency and speeds up computation.
In [ ]:
mesh_shape = [128, 128, 128]
box_size = [128., 128., 128.]
snapshots = jnp.array([0.5, 1.0])

@jax.jit
def run_simulation(omega_c, sigma8):
    # Create a small function to generate the matter power spectrum
    k = jnp.logspace(-4, 1, 128)
    pk = jc.power.linear_matter_power(jc.Planck15(Omega_c=omega_c, sigma8=sigma8), k)
    pk_fn = lambda x: jnp.interp(x.reshape([-1]), k, pk).reshape(x.shape)

    # Create initial conditions
    initial_conditions = linear_field(mesh_shape, box_size, pk_fn, seed=jax.random.PRNGKey(0))

    # Create particles
    cosmo = jc.Planck15(Omega_c=omega_c, sigma8=sigma8)
    
    # Initial displacement
    dx, p, f = lpt(cosmo, initial_conditions, a=0.1,order=1)
    
    # Evolve the simulation forward
    ode_fn = ODETerm(
        make_diffrax_ode(cosmo, mesh_shape, paint_absolute_pos=False))
    solver = LeapfrogMidpoint()

    stepsize_controller = ConstantStepSize()
    res = diffeqsolve(ode_fn,
                      solver,
                      t0=0.1,
                      t1=1.,
                      dt0=0.01,
                      y0=jnp.stack([dx, p], axis=0),
                      args=cosmo,
                      saveat=SaveAt(ts=snapshots),
                      stepsize_controller=stepsize_controller)

    ode_solutions = [sol[0] for sol in res.ys]
    return initial_conditions ,  dx , ode_solutions , res.stats

initial_conditions , lpt_displacements , ode_solutions , solver_stats = run_simulation(0.25, 0.8)
ode_solutions[-1].block_until_ready()
%timeit initial_conditions , lpt_displacements , ode_solutions , solver_stats = run_simulation(0.25, 0.8);ode_solutions[-1].block_until_ready()
print(f"Solver Stats : {solver_stats}")
4.05 s ± 1.54 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
Solver Stats : {'max_steps': Array(4096, dtype=int32, weak_type=True), 'num_accepted_steps': Array(90, dtype=int32, weak_type=True), 'num_rejected_steps': Array(0, dtype=int32, weak_type=True), 'num_steps': Array(90, dtype=int32, weak_type=True)}
In [ ]:
from jaxpm.plotting import plot_fields_single_projection

fields = {"Initial Conditions" : initial_conditions , "LPT Field" : cic_paint_dx(lpt_displacements)}
for i , field in enumerate(ode_solutions):
    fields[f"field_{i}"] = cic_paint_dx(field)
plot_fields_single_projection(fields)
No description has been provided for this image

First and Second Order Lagrangian Perturbation Theory (LPT) Displacements

This section introduces first-order and second-order LPT simulations, controlled by the order argument. First-order LPT captures linear displacements, while second-order LPT includes nonlinear corrections, allowing more accurate modeling of structure formation.

In [ ]:
from functools import partial 

mesh_shape = [128, 128, 128]
box_size = [128., 128., 128.]
snapshots = jnp.array([0.5,1.])

@partial(jax.jit , static_argnums=(2,))
def lpt_simulation(omega_c, sigma8, order=1):
    # Create a small function to generate the matter power spectrum
    k = jnp.logspace(-4, 1, 128)
    pk = jc.power.linear_matter_power(jc.Planck15(Omega_c=omega_c, sigma8=sigma8), k)
    pk_fn = lambda x: jnp.interp(x.reshape([-1]), k, pk).reshape(x.shape)

    # Create initial conditions
    initial_conditions = linear_field(mesh_shape, box_size, pk_fn, seed=jax.random.PRNGKey(0))

    # Create particles
    cosmo = jc.Planck15(Omega_c=omega_c, sigma8=sigma8)
    
    # Initial displacement
    dx, p, f = lpt(cosmo, initial_conditions, a=0.8,order=order)

    return initial_conditions ,  dx

initial_conditions_1 , lpt_displacements_1 = lpt_simulation(0.25, 0.8 , order=1)
lpt_displacements_1.block_until_ready()
initial_conditions_2 , lpt_displacements_2 = lpt_simulation(0.25, 0.8 , order=2)
lpt_displacements_2.block_until_ready()
%timeit initial_conditions_1 , lpt_displacements_1 = lpt_simulation(0.25, 0.8 , order=1);lpt_displacements_1.block_until_ready()
%timeit initial_conditions_2 , lpt_displacements_2 = lpt_simulation(0.25, 0.8, order=2);lpt_displacements_2.block_until_ready()
32.3 ms ± 9.42 μs per loop (mean ± std. dev. of 7 runs, 10 loops each)
42.8 ms ± 9.74 μs per loop (mean ± std. dev. of 7 runs, 10 loops each)
In [5]:
lpt_fields = {"First Order" : cic_paint_dx(lpt_displacements_1) , "Second Order" : cic_paint_dx(lpt_displacements_2)}
plot_fields_single_projection(lpt_fields)
No description has been provided for this image

Custom ODE Solver with Absolute Positions

Just like in the introduction notebook, this example uses absolute particle positions initialized on a uniform grid. We evolve these absolute positions forward using a Cloud-in-Cell (CIC) scheme, which enables clear tracking of particle movement across the simulation volume.

Here, we integrate over multiple snapshots with diffeqsolve and a Leapfrog solver.

In [ ]:
mesh_shape = [128, 128, 128]
box_size = [128., 128., 128.]
snapshots = jnp.array([0.1 ,0.5, 1.])

@jax.jit
def run_simulation(omega_c, sigma8):
    # Create a small function to generate the matter power spectrum
    k = jnp.logspace(-4, 1, 128)
    pk = jc.power.linear_matter_power(jc.Planck15(Omega_c=omega_c, sigma8=sigma8), k)
    pk_fn = lambda x: jnp.interp(x.reshape([-1]), k, pk).reshape(x.shape)

    # Create initial conditions
    initial_conditions = linear_field(mesh_shape, box_size, pk_fn, seed=jax.random.PRNGKey(0))

    # Create particles
    cosmo = jc.Planck15(Omega_c=omega_c, sigma8=sigma8)
    
    particles = uniform_particles(mesh_shape)
    # Initial displacement
    dx, p, f = lpt(cosmo, initial_conditions,particles=particles,a=0.1,order=2)
    
    # Evolve the simulation forward
    ode_fn = ODETerm(
        make_diffrax_ode(cosmo, mesh_shape))
    solver = LeapfrogMidpoint()

    stepsize_controller = ConstantStepSize()
    res = diffeqsolve(ode_fn,
                      solver,
                      t0=0.1,
                      t1=1.,
                      dt0=0.2,
                      y0=jnp.stack([particles + dx, p], axis=0),
                      args=cosmo,
                      saveat=SaveAt(ts=snapshots),
                      stepsize_controller=stepsize_controller)

    ode_particles = [sol[0] for sol in res.ys]
    return initial_conditions ,  particles + dx , ode_particles , res.stats

initial_conditions , lpt_particles , ode_particles , solver_stats = run_simulation(0.25, 0.8)
print(f"Solver Stats : {solver_stats}")
Solver Stats : {'max_steps': Array(4096, dtype=int32, weak_type=True), 'num_accepted_steps': Array(5, dtype=int32, weak_type=True), 'num_rejected_steps': Array(0, dtype=int32, weak_type=True), 'num_steps': Array(5, dtype=int32, weak_type=True)}
In [ ]:
from jaxpm.plotting import plot_fields_single_projection

fields = {"Initial Conditions" : initial_conditions , "LPT Field" : cic_paint(jnp.zeros(mesh_shape) ,lpt_particles)}
for i , field in enumerate(ode_particles[1:]):
    fields[f"field_{i}"] = cic_paint(jnp.zeros(mesh_shape) , field)
plot_fields_single_projection(fields)
No description has been provided for this image

Weighted Field Projection for Central Region

In this cell, we apply custom weights to enhance density specifically in the central 3D region of the grid. By updating weights in this area, we multiply density by a factor of 3, emphasizing the structure in the center of the simulation volume.

We compare:

  • Weighted: Density increased in the central region.
  • Unweighted: Standard CIC painting without additional weighting.
In [ ]:
from jaxpm.plotting import plot_fields_single_projection

center = slice(mesh_shape[0] // 4, 3 * mesh_shape[0] // 4 )
center3d = (slice(None) , center,center) 
weights = jnp.ones_like(initial_conditions)
weights = weights.at[center3d].multiply(3)

weighted = cic_paint_dx(ode_solutions[0], weight=weights)
unweighted = cic_paint_dx(ode_solutions[0] , weight=1.0)

plot_fields_single_projection({"Weighted" : weighted , "Unweighted" : unweighted} , project_axis=0)
No description has been provided for this image

Weighted Field Projection with Absolute Positions

For simulations with absolute positions, we apply a weight factor of 1.3 to the central 3D region. Unlike previous cases using displacements, here the weight affects the absolute particle positions directly, impacting the overall density field differently.

Note: Since the weights apply to absolute positions (not displacements), the result differs, affecting the particle density distribution directly.

In [ ]:
from jaxpm.plotting import plot_fields_single_projection

center = slice(mesh_shape[0] // 4, 3 * mesh_shape[0] // 4 )
center3d = (slice(None) , center,center)  
weights = jnp.ones_like(initial_conditions)
weights = weights.at[center3d].multiply(1.3)

weighted = cic_paint(jnp.zeros(mesh_shape),ode_particles[0], weight=weights)
unweighted = cic_paint(jnp.zeros(mesh_shape),ode_particles[0] , weight=2.0)

plot_fields_single_projection({"Weighted" : weighted , "Unweighted" : unweighted} , project_axis=0)
No description has been provided for this image