Architecture

How the types and data flow are organized.

Types

MaxwellProblem

Specifies the physics: which PDEs to solve.

struct MaxwellProblem
    env::Environment              # materials
    inputs::Vector{FieldConfig}   # pump wavelengths
    outputs::Vector{FieldConfig}  # emission wavelengths (or same as inputs)
    α_loss::Float64               # absorption weighting
end

Environment holds material properties:

Environment(mat_design="Ag", mat_substrate="Ag", mat_fluid=1.33)

FieldConfig describes a single solve:

FieldConfig(λ=532.0, θ=0.0, pol=:y, weight=1.0)

ObjectiveFunction

Abstract type. Implementations must provide three methods:

compute_objective(obj, pde, fields, pt, sim)          # g(E)
compute_adjoint_sources(obj, pde, fields, pt, sim)    # dg/dE
explicit_sensitivity(obj, pde, fields, pf, pt, sim, control)  # dg/dpf

SERSObjective implements this for the SO(3)-averaged trace formula:

SERSObjective(αp=I(3), volume=true, surface=false, use_damage_model=false)

OptimizationProblem

Ties everything together:

mutable struct OptimizationProblem{PDE,Obj}
    pde::MaxwellProblem
    objective::ObjectiveFunction
    sim::Simulation
    foundry_mode::Bool
    pool::SolverCachePool
    control::Control
    p::Vector{Float64}      # design parameters
    g::Float64              # current objective
    grad::Vector{Float64}   # current gradient
end

Optimization flow

p (design params)
    |
    v  filter_grid (foundry) / filter_helmholtz! (3D)
pf (filtered)
    |
    v  interpolate grid->mesh (foundry only)
    v  project_ssp
pt (projected, near-binary)
    |
    v  solve_forward! (Maxwell for each FieldConfig)
E_fields
    |
    |-->  compute_objective -> g
    |
    v  compute_adjoint_sources -> b_adj
    |
    v  solve_adjoint! (reuses the LU factors)
lambda_fields
    |
    v  pde_sensitivity + explicit_sensitivity
dg/dpf
    |
    v  filter_adjoint
dg/dp (gradient)

Key functions

Writing a new objective

1. Create Objectives/MyObjective.jl:

struct MyObjective <: ObjectiveFunction
    # params
end

compute_objective(obj::MyObjective, pde, fields, pt, sim) = ...
compute_adjoint_sources(obj::MyObjective, pde, fields, pt, sim) = ...
explicit_sensitivity(obj::MyObjective, pde, fields, pf, pt, sim, control) = ...

1. Include it in DistributedEmitterOpt.jl and export.

2. Use it:

prob = OptimizationProblem(pde, MyObjective(...), sim, solver)

File organization