|
|
from warnings import warn |
|
|
|
|
|
import numpy as np |
|
|
|
|
|
warn( |
|
|
"DEPRECATED: The file pdeinvbench/losses/fluids_numpy.py is being imported. This is most certainly a bug!" |
|
|
) |
|
|
|
|
|
|
|
|
def _maybe_unsqueeze_3d( |
|
|
u, |
|
|
): |
|
|
""" |
|
|
Given a tensor, makes sure that the last dimension is 1 (channel dim). |
|
|
Helps to ensure number of channels consistency. |
|
|
NOTE: This should only be used within this file. Assumes that u is an unbatched fluid field. |
|
|
Also always assumes well-formed input. |
|
|
""" |
|
|
return u |
|
|
|
|
|
|
|
|
def _maybe_unsqueeze_2d(u): |
|
|
""" |
|
|
Same as 3d version but assumes a tensor of shape (nx, ny, 1) |
|
|
""" |
|
|
return u |
|
|
|
|
|
|
|
|
def _compute_stream_function(vorticity, x, y, fourier=False): |
|
|
""" |
|
|
from the vorticity, compute the stream function and returns in physical space |
|
|
vorticity shape: (nt, nx, ny, 1) Channels is squeezed off |
|
|
""" |
|
|
w = vorticity |
|
|
w = np.squeeze(w) |
|
|
what = np.fft.fft2(w) |
|
|
nx, ny = w.shape[-2], w.shape[-1] |
|
|
dx = x[1] - x[0] |
|
|
dy = y[1] - y[0] |
|
|
kx = np.fft.fftfreq(nx, d=dx) * 2 * np.pi |
|
|
ky = np.fft.fftfreq(ny, d=dy) * 2 * np.pi |
|
|
kx, ky = np.meshgrid(kx, ky) |
|
|
wavenumbers_squared = kx**2 + ky**2 |
|
|
|
|
|
psi_hat = np.zeros_like(what, dtype=np.complex128) |
|
|
psi_hat[wavenumbers_squared > 0] = what[wavenumbers_squared > 0] / ( |
|
|
-wavenumbers_squared[wavenumbers_squared > 0] |
|
|
) |
|
|
|
|
|
if fourier: |
|
|
return _maybe_unsqueeze_2d(psi_hat) |
|
|
else: |
|
|
return np.fft.ifft2(psi_hat).real |
|
|
|
|
|
|
|
|
def compute_stream_function(vorticity, x, y, fourier=False): |
|
|
nt = vorticity.shape[0] |
|
|
out = [] |
|
|
for i in range(nt): |
|
|
out.append(_compute_stream_function(vorticity[i], x, y, fourier=fourier)) |
|
|
return np.asarray(out) |
|
|
|
|
|
|
|
|
def _compute_stream_function_jaxcfd(vorticity, x, y, fourier=False): |
|
|
w = vorticity |
|
|
w = np.squeeze(w) |
|
|
what = np.fft.fft2(w) |
|
|
nx, ny = w.shape[-2], w.shape[-1] |
|
|
dx = x[1] - x[0] |
|
|
dy = y[1] - y[0] |
|
|
kx = np.fft.fftfreq(nx, d=dx) * 2 * np.pi |
|
|
ky = np.fft.fftfreq(ny, d=dy) * 2 * np.pi |
|
|
kx, ky = np.meshgrid(kx, ky) |
|
|
wavenumbers_squared = kx**2 + ky**2 |
|
|
|
|
|
|
|
|
lap = wavenumbers_squared |
|
|
lap[0, 0] = 1 |
|
|
psi_hat = -1 / lap * what |
|
|
if fourier: |
|
|
return psi_hat |
|
|
return np.fft.ifft2(psi_hat).real |
|
|
|
|
|
|
|
|
def compute_stream_function_jaxcfd(vorticity, x, y, fourier=False): |
|
|
nt = vorticity.shape[0] |
|
|
out = [] |
|
|
for i in range(nt): |
|
|
out.append(_compute_stream_function_jaxcfd(vorticity[i], x, y, fourier=fourier)) |
|
|
return np.asarray(out) |
|
|
|
|
|
|
|
|
def advection(vorticity, x, y, stream_func): |
|
|
psi_hat = stream_func(vorticity, x, y, fourier=True) |
|
|
w = vorticity |
|
|
w = np.squeeze(w) |
|
|
|
|
|
nx, ny = w.shape[-2], w.shape[-1] |
|
|
dx = x[1] - x[0] |
|
|
dy = y[1] - y[0] |
|
|
kx = np.fft.fftfreq(nx, d=dx) * 2 * np.pi |
|
|
ky = np.fft.fftfreq(ny, d=dy) * 2 * np.pi |
|
|
kx, ky = np.meshgrid(kx, ky) |
|
|
vx_hat = 1j * ky * psi_hat |
|
|
vy_hat = -1j * kx * psi_hat |
|
|
vx = np.fft.ifft2(vx_hat).real |
|
|
vy = np.fft.ifft2(vy_hat).real |
|
|
w_dx = np.gradient(w, dx, axis=1) |
|
|
w_dy = np.gradient(w, dy, axis=2) |
|
|
|
|
|
adv = vx * w_dx + vy * w_dy |
|
|
return _maybe_unsqueeze_3d(adv), _maybe_unsqueeze_3d(vx), _maybe_unsqueeze_3d(vy) |
|
|
|
|
|
|
|
|
def second_order_gradient(field, d_coord, axis): |
|
|
return np.gradient(np.gradient(field, d_coord, axis=axis), d_coord, axis=axis) |
|
|
|
|
|
|
|
|
def laplacian(vorticity, x, y): |
|
|
w = vorticity |
|
|
w = np.squeeze(w) |
|
|
nx, ny = w.shape[-2], w.shape[-1] |
|
|
dx = x[1] - x[0] |
|
|
dy = y[1] - y[0] |
|
|
w_dxx = second_order_gradient(w, dx, axis=1) |
|
|
w_dyy = second_order_gradient(w, dy, axis=2) |
|
|
return w_dxx + w_dyy |
|
|
|
|
|
|
|
|
def tf_residual_numpy(w, t, x, y, nu): |
|
|
alpha = 0.1 |
|
|
forced_mode = 2 |
|
|
dy = y[1] - y[0] |
|
|
dx = x[1] - x[0] |
|
|
f = forced_mode * np.cos(forced_mode * y) |
|
|
|
|
|
f = np.reshape(f, (1, 1, -1, 1)) |
|
|
f = np.broadcast_to(f, w.shape) |
|
|
|
|
|
damping = alpha * w |
|
|
|
|
|
dt: float = t[1] - t[0] |
|
|
dwdt = np.gradient(w, dt, axis=0) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
psi = compute_stream_function(w, x, y) |
|
|
psi = np.expand_dims(psi, axis=-1) |
|
|
|
|
|
|
|
|
|
|
|
adv, vx, vy = advection(w, x, y, compute_stream_function) |
|
|
prep_plot = lambda a: np.expand_dims(a, axis=-1) |
|
|
advection_terms = {"adv": adv, "vx": vx, "vy": vy} |
|
|
|
|
|
lap = laplacian(w, x, y) |
|
|
|
|
|
|
|
|
diffusion = nu * lap |
|
|
|
|
|
|
|
|
|
|
|
adv = np.expand_dims(adv, axis=-1) |
|
|
diffusion = np.expand_dims(diffusion, axis=-1) |
|
|
|
|
|
residual = -1 * dwdt + -1 * adv + diffusion - damping + f |
|
|
return residual |
|
|
|