Failed nonlinear solves get recorded on tape even though no value is recorded

[Gabriel Balaban]

Sometimes optimization algorithms can go into areas of parameter space where a nonlinear solver crashes. If one has no a-priori guarantee of convergence then the best one can do is tune some parameters and try again. Unfortunately the failed solve attempts get recorded by dolfin-adjoint, as the script below demonstrates.

from dolfin import *
from dolfin_adjoint import *

mesh = UnitSquareMesh(3,3)
V = FunctionSpace(mesh, "CG",1)
u = Function(V)
v = TestFunction(V)

alpha = Constant(1.0e10)

F = inner(grad(u), grad(v))*dx + alpha*u**2*v*dx

bc = DirichletBC(V, Expression("x[0]"), "on_boundary")

try:
    #I am going to fail
    solve(F == 0, u, bc)
except:
    #Oops alpha was too big better try something smaller.
    alpha.assign(1.0)
    #Now I can solve the problem!
    solve(F == 0, u, bc)

I = Functional(u*dx)
Rd = ReducedFunctional(I, Control(alpha))
#I get confused by the solve that failed.
Rd.derivative()