-
Notifications
You must be signed in to change notification settings - Fork 18
/
solve.m
221 lines (179 loc) · 7.65 KB
/
solve.m
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
function [eps] = solve(spec, max_iters, min_grad, varargin)
% EPS = SOLVE(SPEC, MAX_ITERS, MIN_GRAD, [METHOD])
%
% Description
% Perform an objective-first optimization for a nanophotonic waveguide
% coupler.
%
% Inputs
% SPEC: Structure.
% The output of SETUP(). SPEC defines the waveguide coupler problem.
%
% MAX_ITERS: Non-negative integer.
% Maximum number of iterations for which to run the optimization.
%
% MIN_GRAD: Non-negative scalar.
% Minimum value for the norm of the gradient. The optimization will
% stop once the norm of the gradient is below MIN_GRAD.
%
% Note that the norm of the gradient is used, and not the norm-squared.
%
% METHOD: Character string (optional).
% The name of the numerical method to use to solve the problem.
% Currently only the 'alt_dir' option is supported. 'alt_dir'
% requires the CVX package (www.stanford.edu/~boyd/cvx).
%
% Outputs
% EPS: 2d array.
% The permittivity values of the final design.
% Implementation notes
% * We attempt to completely use linear-algebra language in this function,
% which means everything is either a matrix or a vector (no 2-d arrays
% for example). The implementation is much clearer this way.
% * We fix the range of p to be from 0 to 1 (inclusive). In order to match
% the desired range in epsilon values, p is scaled and offset.
% * For numerical reasons (convexity in the p variable) we use the inverse
% of epsilon instead of epsilon itself.
% Make sure we have access to the cvx package.
path(path, genpath('cvx'));
dims = size(spec.eps0);
N = prod(dims);
%
% Form relevant matrices, and get initial values of x and p.
%
[A, S] = ob1_matrices(dims); % Relevant matrices.
x0 = spec.Hz0(:); % Initial value of x.
x_int = S.int * x0; % Interior values of the field (variable).
x_bnd = S.bnd * x0; % Boundary values of the field (fixed).
p_scale = diff(spec.eps_lims.^-1); % Scaling factor for p.
p_offset = (spec.eps_lims(1).^-1) * ones(N,1); % Offset for values of p.
p0 = p_scale^-1 * (spec.eps0(:).^-1 - p_offset); % Initial value of p.
p_int = S.int * p0; % Interior values of the structure (variable).
p_bnd = S.bnd * p0; % Boundary values of the structure (fixed).
%
% Construct helper functions related to the physics residual.
%
my_diag = @(z) spdiags(z, 0, length(z), length(z)); % Make diagonal matrix.
p2ie = @(p) p_scale * p + p_offset; % Transform from p to 1/eps.
% Physics operator in terms of x.
A_x = @(p) A{1} * my_diag(A{2} * p2ie(p)) * A{3} - spec.omega^2 * speye(N);
% Physics operator in terms of p (A_p * p - b_p = 0).
A_p = @(x) p_scale * A{1} * my_diag(A{3} * x) * A{2};
b_p = @(x) spec.omega^2 * x - A_p(x) * (p_offset ./ p_scale);
% Physics residual.
phys_res = @(x, p) norm(S.res * (A_x(p) * x))^2;
% phys_res_alt = @(x, p) norm(S.res * (A_p(x) * p - b_p(x)))^2;
% Gradient of the physics residual.
grad = @(x, p) [((S.res * A_x(p))' * (S.res * A_x(p)) * x); ...
(S.res * A_p(x))'*(S.res * (A_p(x) * p - b_p(x)))];
%
% Helper functions for optimization process.
%
% Functions to include add boundary values to interior values of x and p.
p_full = @(pin) S.int' * pin + S.bnd' * p_bnd;
x_full = @(xin) S.int' * xin + S.bnd' * x_bnd;
% Progress functions.
progress = @(xin, pin) [phys_res(x_full(xin), p_full(pin)), ...
norm(grad(x_full(xin), p_full(pin)))];
print_prog = @(iter, xin, pin) ...
fprintf('%5d: \t%1.5e \t%1.5e\n', iter, progress(xin, pin));
print_header = @(str) ...
fprintf('%s iter#: \t[phys_res] \t[grad_norm]\n', str);
%
% Perform the optimization.
%
% Resolve what optimization method we should use.
if isempty(varargin)
method = 'alt_dir';
else
method = varargin{1};
end
switch method
case 'alt_dir'
% Alternating directions method.
% * The basic idea is to alternately optimize for x and then p.
% * This allows us to use standard, guaranteed-to-work solvers.
% * Although this method is slow, it should always approach a
% solution.
fprintf('Starting the alternating directions solver...\n');
print_header('x/p'); % Print header information for progress.
start_time = tic; % For timing purposes.
for k = 1 : max_iters
% Solve for x_int.
x_int = (A_x(p_full(p_int)) * S.int') \ ...
(-A_x(p_full(p_int)) * S.bnd' * x_bnd);
fprintf('(x) '); print_prog(k, x_int, p_int)
% Solve for p_int.
cvx_quiet(true);
cvx_begin
variable p_int(length(p_int))
minimize norm(A_p(x_full(x_int)) * ...
(S.int' * p_int + S.bnd' * p_bnd) - ...
b_p(x_full(x_int)))
subject to
p_int >= 0
p_int <= 1
cvx_end
fprintf('(p) '); print_prog(k, x_int, p_int)
% Visualize.
ob1_plot(dims, {'p', p_full(p_int)}, ...
{'|x|', abs(x_full(x_int))}, ...
{'Re(x)', real(x_full(x_int))});
% Check for gradient norm stopping condition.
if (norm(grad(x_full(x_int), p_full(p_int))) < min_grad)
fprintf('Gradient norm stopping condition satisfied.');
break
end
end
case 'alt_dir_mod'
% Alternating directions method.
% * The basic idea is to alternately optimize for x and then p.
% * This allows us to use standard, guaranteed-to-work solvers.
% * Although this method is slow, it should always approach a
% solution.
fprintf('Starting the alternating directions solver...\n');
print_header('x/p'); % Print header information for progress.
start_time = tic; % For timing purposes.
ptot = norm(p_int, 1);
for k = 1 : max_iters
% Solve for x_int.
x_int = (A_x(p_full(p_int)) * S.int') \ ...
(-A_x(p_full(p_int)) * S.bnd' * x_bnd);
fprintf('(x) '); print_prog(k, x_int, p_int)
% Solve for p_int.
cvx_quiet(true);
cvx_begin
variable p_int(length(p_int))
minimize norm(A_p(x_full(x_int)) * ...
(S.int' * p_int + S.bnd' * p_bnd) - ...
b_p(x_full(x_int)))
subject to
p_int >= 0
p_int <= 1
norm(p_int, 1) <= ptot
cvx_end
fprintf('(p) '); print_prog(k, x_int, p_int)
% Visualize.
ob1_plot(dims, {'p', p_full(p_int)}, ...
{'|x|', abs(x_full(x_int))}, ...
{'Re(x)', real(x_full(x_int))});
% Check for gradient norm stopping condition.
if (norm(grad(x_full(x_int), p_full(p_int))) < min_grad)
fprintf('Gradient norm stopping condition satisfied.');
break
end
end
otherwise
error('Invalid choice of METHOD (%s).', method);
end
%
% Extract the full design and print final result.
%
eps = reshape(p2ie(p_full(p_int)).^-1, dims); % Get full structure.
fprintf('\n%d iterations completed in %1.1f seconds.\n', k, toc(start_time));
fprintf('Final physics residual: %1.5e\nFinal gradient norm: %1.5e\n', ...
progress(x_int, p_int));
% Visualize.
ob1_plot(dims, {'p', p_full(p_int)}, ...
{'|x|', abs(x_full(x_int))}, ...
{'Re(x)', real(x_full(x_int))});