File size: 12,945 Bytes
357baae
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
"""Exact piecewise acoustic modes for a planar unequal-speed interface.

The finite continuous modal initial field is fixed before grid projection.
Continuous roots, full discrete modes, and reduced modes are constructed
independently. The separated solvers evaluate modal responses without time stepping.
"""
from functools import lru_cache
from pathlib import Path
import json
import numpy as np
from numpy.polynomial.legendre import leggauss
from scipy import linalg, sparse, special
from scipy.optimize import brentq

HERE = Path(__file__).resolve().parent
CFG = json.loads((HERE / "config.json").read_text())


def gauss_interval(order, a, b):
    x, w = leggauss(order)
    return (a+b)/2 + (b-a)*x/2, (b-a)*w/2


def cosine_q(q2, d):
    q2, d = np.broadcast_arrays(np.asarray(q2, dtype=float), np.asarray(d, dtype=float))
    ans = np.empty(q2.shape)
    positive = q2 >= 0
    ans[positive] = np.cos(np.sqrt(q2[positive]) * d[positive])
    ans[~positive] = np.cosh(np.sqrt(-q2[~positive]) * d[~positive])
    return ans


def q_sine(q2, d):
    q2, d = np.broadcast_arrays(np.asarray(q2, dtype=float), np.asarray(d, dtype=float))
    ans = np.empty(q2.shape)
    positive = q2 >= 0
    qq = np.sqrt(q2[positive])
    ans[positive] = qq * np.sin(qq * d[positive])
    qq = np.sqrt(-q2[~positive])
    ans[~positive] = -qq * np.sinh(qq * d[~positive])
    return ans


def mean_cosine(q2, centre, width):
    q2, centre = np.broadcast_arrays(np.asarray(q2, dtype=float), np.asarray(centre, dtype=float))
    answer = cosine_q(q2, centre)
    positive = q2 >= 0
    answer[positive] *= np.sinc(np.sqrt(q2[positive]) * width / (2*np.pi))
    z = np.sqrt(-q2[~positive]) * width/2
    answer[~positive] *= np.sinh(z)/z
    return answer


def integral_cosine_squared(q2, length):
    if q2 >= 0:
        return length/2 * (1 + np.sinc(2*np.sqrt(q2)*length/np.pi))
    z = 2*np.sqrt(-q2)*length
    return length/2 * (1 + np.sinh(z)/z)


def material_parameters(density=None, bulk=None):
    rho = np.asarray(CFG["density"] if density is None else density, dtype=float)
    kk = np.asarray(CFG["bulk_modulus"] if bulk is None else bulk, dtype=float)
    return rho, kk, kk/rho


def characteristic(omega, ell, density=None, bulk=None, interface=None):
    rho, kk, c2 = material_parameters(density, bulk)
    a = CFG["interface_x"] if interface is None else interface
    omega = np.asarray(omega)
    qq = omega[..., None]**2/c2 - (ell*np.pi)**2
    cc = cosine_q(qq, [a, 1-a])
    ss = q_sine(qq, [a, 1-a])/rho
    return cc[..., 0]*ss[..., 1] + cc[..., 1]*ss[..., 0]


@lru_cache(maxsize=256)
def roots(ell, count, density=None, bulk=None, scan_step=None):
    rho, kk, c2 = material_parameters(density, bulk)
    step = CFG["root_scan_step"] if scan_step is None else scan_step
    lower = np.sqrt(c2.min()) * ell*np.pi
    upper = np.sqrt(c2.max()) * np.pi*np.sqrt((count+2)**2 + ell**2)
    exact_lowest = ell == 0 or np.isclose(c2[0], c2[1], rtol=1e-14, atol=0.)
    first = lower+1e-8 if exact_lowest else lower
    grid = np.linspace(first, upper, int(np.ceil((upper-lower)/step))+1)
    value = characteristic(grid, ell, tuple(rho), tuple(kk))
    indices = np.flatnonzero(np.signbit(value[:-1]) != np.signbit(value[1:]))
    found = [float(lower)] if exact_lowest else []
    for j in indices:
        root = brentq(characteristic, grid[j], grid[j+1],
                      args=(ell, tuple(rho), tuple(kk)), xtol=5e-13, rtol=2e-14)
        if not found or root-found[-1] > 1e-7:
            found.append(root)
        if len(found) == count:
            break
    if len(found) != count:
        raise RuntimeError(f"Root count failed for ell={ell}: {len(found)} instead of {count}")
    return np.asarray(found)


def mode_parameters(omega, ell, density=None, bulk=None):
    rho, kk, c2 = material_parameters(density, bulk)
    a = CFG["interface_x"]
    q2 = omega**2/c2 - (ell*np.pi)**2
    cc = cosine_q(q2, [a, 1-a])
    ss = q_sine(q2, [a, 1-a])/rho
    matrix = np.array([[cc[0], -cc[1]], [ss[0], ss[1]]])
    # Scale the flux equation by a physical wavenumber, rather than normalizing
    # each row to unit norm: both sine values vanish at some exact roots, and
    # amplifying that roundoff-sized row would corrupt the pressure condition.
    flux_scale = max(1., np.sqrt(omega**2/c2.min()+(ell*np.pi)**2)/rho.min())
    scaled = matrix.copy()
    scaled[1] /= flux_scale
    _, _, vh = linalg.svd(scaled, check_finite=False)
    amplitude = vh[-1]
    if amplitude[0] < 0:
        amplitude = -amplitude
    norm2 = sum(amplitude[i]**2/kk[i] * integral_cosine_squared(q2[i], d)
                for i, d in enumerate((a, 1-a)))
    amplitude /= np.sqrt(norm2)
    return q2, amplitude


def x_values(x, q2, amplitude, average_width=None):
    x = np.asarray(x)
    left = x < CFG["interface_x"]
    ans = np.empty((len(x), len(q2)))
    for side, active, distance in ((0, left, x[left]), (1, ~left, 1-x[~left])):
        if average_width is None:
            vals = cosine_q(q2[None, :, side], distance[:, None])
        else:
            vals = mean_cosine(q2[None, :, side], distance[:, None], average_width)
        ans[active] = vals * amplitude[None, :, side]
    return ans


def y_values(y, maximum):
    ll = np.arange(maximum+1)
    return np.cos(np.pi*np.asarray(y)[:, None]*ll) * np.where(ll == 0, 1., np.sqrt(2.))


def gaussian_integrals(edges, centre):
    z = (np.asarray(edges)-centre)/(np.sqrt(2.)*CFG["sigma"])
    return np.sqrt(np.pi/2)*CFG["sigma"]*np.diff(special.erf(z))


@lru_cache(maxsize=8)
def continuous_data(order=None, scan_step=None):
    order = CFG["quadrature_order"] if order is None else order
    rho, kk, c2 = material_parameters()
    nx = CFG["max_x_mode"]+1
    nl = CFG["max_y_mode"]+1
    omega = np.stack([roots(l, nx, scan_step=scan_step) for l in range(nl)])
    q2 = np.empty((nl, nx, 2))
    amplitudes = np.empty_like(q2)
    for l in range(nl):
        for m in range(nx):
            q2[l, m], amplitudes[l, m] = mode_parameters(omega[l, m], l)
    a = CFG["interface_x"]
    xl, wl = gauss_interval(order, 0., a)
    xr, wr = gauss_interval(order, a, 1.)
    x, wx = np.r_[xl, xr], np.r_[wl, wr]
    beta = np.r_[np.full(order, 1/kk[0]), np.full(order, 1/kk[1])]
    y, wy = gauss_interval(order, 0., 1.)
    xx = np.stack([x_values(x, q2[l], amplitudes[l]) for l in range(nl)])
    yy = y_values(y, nl-1)
    coefficients = np.empty((nl, nx, 3))
    receiver_modes = np.empty((nl, nx, 4))
    gram = np.stack([xx[l].T @ ((wx*beta)[:, None]*xx[l]) for l in range(nl)])
    for s, (cx, cy) in enumerate(CFG["source_centres"]):
        gx = np.exp(-.5*((x-cx)/CFG["sigma"])**2)
        gy = np.exp(-.5*((y-cy)/CFG["sigma"])**2)
        sx = np.einsum("lim,i->lm", xx, wx*beta*gx)
        sy = yy.T @ (wy*gy)
        coeff = sx*sy[:, None]
        coeff[0, 0] = 0.
        coefficients[:, :, s] = coeff*np.sqrt(2*CFG["initial_energy"])/np.linalg.norm(coeff)
    for r, (cx, cy) in enumerate(CFG["receiver_centres"]):
        gx = np.exp(-.5*((x-cx)/CFG["sigma"])**2)
        gy = np.exp(-.5*((y-cy)/CFG["sigma"])**2)
        rx = np.einsum("lim,i->lm", xx, wx*gx)/gaussian_integrals([0., 1.], cx)[0]
        ry = (yy.T @ (wy*gy))/gaussian_integrals([0., 1.], cy)[0]
        receiver_modes[:, :, r] = rx*ry[:, None]
    return {"omega": omega, "q_squared": q2, "amplitudes": amplitudes,
            "coefficients": coefficients, "receiver_modes": receiver_modes,
            "x_gram": gram, "y_gram": yy.T @ (wy[:, None]*yy)}


def exact_receiver(times, order=None, scan_step=None):
    data = continuous_data(order, scan_step)
    phase = np.cos(np.asarray(times)[:, None, None]*data["omega"][None])
    return np.einsum("tlm,lms,lmr->trs", phase, data["coefficients"],
                     data["receiver_modes"], optimize=True)


def exact_pressure_norm_squared(times):
    data = continuous_data()
    phase = np.cos(np.asarray(times)[:, None, None]*data["omega"][None])
    return np.einsum("tlm,lms->ts", phase**2, data["coefficients"]**2, optimize=True)


@lru_cache(maxsize=8)
def averaged_modes(n):
    if n % 2:
        raise ValueError("The interface requires an even grid")
    data = continuous_data()
    x = (np.arange(n)+.5)/n
    xx = np.stack([x_values(x, data["q_squared"][l], data["amplitudes"][l], 1/n)
                   for l in range(CFG["max_y_mode"]+1)])
    ll = np.arange(CFG["max_y_mode"]+1)
    yy = y_values(x, CFG["max_y_mode"])*np.sinc(ll/(2*n))[None, :]
    return xx, yy


def mass_vector(n):
    x = (np.arange(n)+.5)/n
    bulk = np.where(x < CFG["interface_x"], *CFG["bulk_modulus"])
    return np.repeat(1/bulk, n)/n**2


def synthesize(x_modes, y_modes, omega, coefficients, times):
    """Return physical or normalized pressure for factorized modal vectors.

    x_modes has shape (l,i,m), y_modes (j,l), omega (l,m).
    """
    times = np.asarray(times)
    modal = np.cos(times[:, None, None]*omega[None])[:, :, :, None]*coefficients[None]
    nt, nl, nm, ns = modal.shape
    temporary = x_modes @ modal.transpose(1, 2, 0, 3).reshape(nl, nm, nt*ns)
    temporary = temporary.reshape(nl, len(x_modes[0]), nt, ns).transpose(2, 1, 0, 3)
    return np.einsum("tils,jl->tijs", temporary, y_modes, optimize=True).reshape(nt, -1, ns)


def exact_mass_weighted_pressure(n, times):
    data = continuous_data()
    xx, yy = averaged_modes(n)
    return synthesize(xx, yy, data["omega"], data["coefficients"], times)*np.sqrt(mass_vector(n))[None, :, None]


def source_receiver(n, mass):
    B = exact_mass_weighted_pressure(n, [0.])[0]
    edges = np.linspace(0., 1., n+1)
    windows = []
    for cx, cy in CFG["receiver_centres"]:
        gx, gy = gaussian_integrals(edges, cx), gaussian_integrals(edges, cy)
        windows.append(np.outer(gx/gx.sum(), gy/gy.sum()).ravel()/np.sqrt(mass))
    return B, np.stack(windows, axis=-1)


def assembly(n):
    if n % 2:
        raise ValueError("The interface requires an even grid")
    x = (np.arange(n)+.5)/n
    X, Y = np.meshgrid(x, x, indexing="ij")
    rho = np.where(X < CFG["interface_x"], *CFG["density"])
    bulk = np.where(X < CFG["interface_x"], *CFG["bulk_modulus"])
    idx = np.arange(n*n).reshape(n, n)
    a = np.r_[idx[:-1].ravel(), idx[:, :-1].ravel()]
    b = np.r_[idx[1:].ravel(), idx[:, 1:].ravel()]
    faces = np.arange(len(a))
    face_rho = .5*(rho.ravel()[a]+rho.ravel()[b])
    D = sparse.coo_matrix((np.r_[np.ones(len(a)), -np.ones(len(a))],
                          (np.r_[a, b], np.r_[faces, faces])), shape=(n*n, len(a))).tocsr()
    mass = (1/(n*n*bulk)).ravel()
    C = sparse.diags(1/np.sqrt(mass)) @ D @ sparse.diags(1/np.sqrt(face_rho))
    J = sparse.bmat([[None, -C], [C.T, None]], format="csr")
    B, R = source_receiver(n, mass)
    return mass, (C@C.T).tocsr(), J, B, R, rho, bulk


def x_tridiagonal(n, density=None, bulk=None):
    rho, kk, c2 = material_parameters(density, bulk)
    x = (np.arange(n)+.5)/n
    rr = np.where(x < CFG["interface_x"], *rho)
    bb = np.where(x < CFG["interface_x"], *kk)
    inv = 2/(rr[:-1]+rr[1:])
    diagonal = n*n*bb*np.r_[inv[0], inv[:-1]+inv[1:], inv[-1]]
    off = -n*n*np.sqrt(bb[:-1]*bb[1:])*inv
    return diagonal, off, bb/rr


def separable_setup(n):
    dx, ex, speed2 = x_tridiagonal(n)
    dy = np.r_[1., np.full(n-2, 2.), 1.]*n*n
    ly, Vy = linalg.eigh_tridiagonal(dy, np.full(n-1, -n*n, dtype=float))
    ly[np.abs(ly) < 1e-8] = 0.
    lam = np.empty((n, n))
    Vx = np.empty((n, n, n))
    for ell in range(n):
        lam[ell], Vx[ell] = linalg.eigh_tridiagonal(dx+speed2*ly[ell], ex)
    lam[np.abs(lam) < 1e-8] = 0.
    if np.min(lam) < 0:
        raise RuntimeError("Indefinite separated stiffness")
    mass = mass_vector(n)
    B, R = source_receiver(n, mass)
    By = np.einsum("ijs,jl->lis", B.reshape(n, n, 3), Vy, optimize=True)
    Ry = np.einsum("ijr,jl->lir", R.reshape(n, n, 4), Vy, optimize=True)
    coeff = np.matmul(Vx.transpose(0, 2, 1), By)
    observation = np.matmul(Vx.transpose(0, 2, 1), Ry)
    return {"omega": np.sqrt(lam), "x_eigenvectors": Vx, "y_eigenvectors": Vy,
            "initial_modal_coefficients": coeff, "receiver_modes": observation, "mass": mass}


def separable_receiver(model, times):
    phase = np.cos(np.asarray(times)[:, None, None]*model["omega"][None])
    return np.einsum("tlm,lms,lmr->trs", phase, model["initial_modal_coefficients"],
                     model["receiver_modes"], optimize=True)


def separable_pressure(model, times):
    return synthesize(model["x_eigenvectors"], model["y_eigenvectors"],
                      model["omega"], model["initial_modal_coefficients"], times)


def full_eigenvectors(model):
    n = len(model["mass"])**.5
    n = int(n)
    vectors = np.einsum("lim,jl->ijlm", model["x_eigenvectors"], model["y_eigenvectors"],
                        optimize=True).reshape(n*n, n*n)
    order = np.argsort(model["omega"].ravel())
    return model["omega"].ravel()[order], vectors[:, order], order