Brusselator 1d (automatic)

References

This example is taken from Numerical Bifurcation Analysis of Periodic Solutions of Partial Differential Equations, Lust, 1997.

We look at the Brusselator in 1d. The equations are as follows

\[\begin{aligned} \frac { \partial X } { \partial t } & = \frac { D _ { 1 } } { l ^ { 2 } } \frac { \partial ^ { 2 } X } { \partial z ^ { 2 } } + X ^ { 2 } Y - ( β + 1 ) X + α \\ \frac { \partial Y } { \partial t } & = \frac { D _ { 2 } } { l ^ { 2 } } \frac { \partial ^ { 2 } Y } { \partial z ^ { 2 } } + β X - X ^ { 2 } Y \end{aligned}\]

with Dirichlet boundary conditions

\[\begin{array} { l } { X ( t , z = 0 ) = X ( t , z = 1 ) = α } \\ { Y ( t , z = 0 ) = Y ( t , z = 1 ) = β / α } \end{array}\]

These equations have been introduced to reproduce an oscillating chemical reaction. There is an obvious equilibrium $(α, β / α)$. Here, we consider bifurcations with respect to the parameter $l$.

We start by writing the PDE

using Revise
using BifurcationKit, LinearAlgebra, Plots, SparseArrays, Setfield, Parameters
const BK = BifurcationKit

f1(u, v) = u * u * v
norminf = x -> norm(x, Inf)

function Fbru(x, p)
	@unpack α, β, D1, D2, l = p
	f = similar(x)
	n = div(length(x), 2)
	h = 1.0 / n; h2 = h*h
	c1 = D1 / l^2 / h2
	c2 = D2 / l^2 / h2

	u = @view x[1:n]
	v = @view x[n+1:2n]

	# Dirichlet boundary conditions
	f[1]   = c1 * (α      - 2u[1] + u[2] ) + α - (β + 1) * u[1] + f1(u[1], v[1])
	f[end] = c2 * (v[n-1] - 2v[n] + β / α)			 + β * u[n] - f1(u[n], v[n])

	f[n]   = c1 * (u[n-1] - 2u[n] +  α  )  + α - (β + 1) * u[n] + f1(u[n], v[n])
	f[n+1] = c2 * (β / α  - 2v[1] + v[2])			 + β * u[1] - f1(u[1], v[1])

	for i=2:n-1
		  f[i] = c1 * (u[i-1] - 2u[i] + u[i+1]) + α - (β + 1) * u[i] + f1(u[i], v[i])
		f[n+i] = c2 * (v[i-1] - 2v[i] + v[i+1])			  + β * u[i] - f1(u[i], v[i])
	end
	return f
end

For computing periodic orbits, we will need a Sparse representation of the Jacobian:

function Jbru_sp(x, p)
	@unpack α, β, D1, D2, l = p
	# compute the Jacobian using a sparse representation
	n = div(length(x), 2)
	h = 1.0 / n; h2 = h*h

	c1 = D1 / p.l^2 / h2
	c2 = D2 / p.l^2 / h2

	u = @view x[1:n]
	v = @view x[n+1:2n]

	diag   = zeros(eltype(x), 2n)
	diagp1 = zeros(eltype(x), 2n-1)
	diagm1 = zeros(eltype(x), 2n-1)

	diagpn = zeros(eltype(x), n)
	diagmn = zeros(eltype(x), n)

	@. diagmn = β - 2 * u * v
	@. diagm1[1:n-1] = c1
	@. diagm1[n+1:end] = c2

	@. diag[1:n]    = -2c1 - (β + 1) + 2 * u * v
	@. diag[n+1:2n] = -2c2 - u * u

	@. diagp1[1:n-1] = c1
	@. diagp1[n+1:end] = c2

	@. diagpn = u * u
	return spdiagm(0 => diag, 1 => diagp1, -1 => diagm1, n => diagpn, -n => diagmn)
end
Tip

We could have used DiffEqOperators.jl like for the Swift-Hohenberg tutorial.

Finally, it will prove useful to have access to the hessian and third derivative

# we group the differentials together
jet  = BK.get3Jet(Fbru, Jbru_sp)

We shall now compute the equilibria and their stability.

n = 300

# parameters of the Brusselator model and guess for the stationary solution
par_bru = (α = 2., β = 5.45, D1 = 0.008, D2 = 0.004, l = 0.3)
sol0 = vcat(par_bru.α * ones(n), par_bru.β/par_bru.α * ones(n))

For the eigensolver, we use a Shift-Invert algorithm (see Eigen solvers (Eig))

eigls = EigArpack(1.1, :LM)

We continue the trivial equilibrium to find the Hopf points

opt_newton = NewtonPar(eigsolver = eigls, verbose = false)
opts_br_eq = ContinuationPar(dsmin = 0.001, dsmax = 0.01, ds = 0.001,
	pMax = 1.9, detectBifurcation = 3, nev = 21, plotEveryStep = 50,
	newtonOptions = NewtonPar(eigsolver = eigls, tol = 1e-9), maxSteps = 1060,
	# specific options for precise localization of Hopf points
	nInversion = 6)

	br, = @time continuation(Fbru, Jbru_sp, sol0, par_bru, (@lens _.l),
		opts_br_eq, verbosity = 0,
		plot = true,
		printSolution = (x,p) -> x[n÷2], normC = norminf)

We obtain the following bifurcation diagram with 3 Hopf bifurcation points

Normal form computation

We can compute the normal form of the Hopf points as follows

hopfpt = BK.computeNormalForm(jet..., br, 1)

and you should get

julia> hopfpt
SuperCritical - Hopf bifurcation point at l ≈ 0.5113310149554013.
Period of the periodic orbit ≈ 2.9367552006841753
Normal form z⋅(a⋅δp + b⋅|z|²): 
(a = 0.8799941318427783 + 0.5689746667563035im, b = -0.0015608102901479592 + 0.0015634810970084371im)

Continuation of Hopf points

We use the bifurcation points guesses located in br.specialpoint to turn them into precise bifurcation points. For the second one, we have

# index of the Hopf point in br.specialpoint
ind_hopf = 2
hopfpoint, _, flag = @time newton(Fbru, Jbru_sp,
	br, ind_hopf; normN = norminf)
flag && printstyled(color=:red, "--> We found a Hopf Point at l = ", hopfpoint.p[1], ", ω = ", hopfpoint.p[2], ", from l = ", br.specialpoint[ind_hopf].param, "\n")

which produces

--> We found a Hopf Point at l = 1.0239851696548035, ω = 2.1395092895339842, from l = 1.0353910524340078

We now perform a Hopf continuation with respect to the parameters l, β

Tip

You don't need to call newton first in order to use continuation.

optcdim2 = ContinuationPar(dsmin = 0.001, dsmax = 0.05, ds= 0.01, pMax = 6.5, pMin = 0.0, newtonOptions = opt_newton)
br_hopf, = @time continuation(Fbru, Jbru_sp,
	br, ind_hopf, (@lens _.β),
	optcdim2, verbosity = 2, 
	normC = norminf)

which gives using plot(br_hopf)

Computation of the branch of periodic orbits (Finite differences)

We now compute the bifurcated branches of periodic solutions from the Hopf points using Periodic orbits based on trapezoidal rule. One has just to pass a PeriodicOrbitTrapProblem.

We start by providing a linear solver and some options for the continuation to work

# automatic branch switching from Hopf point
opt_po = NewtonPar(tol = 1e-10, verbose = true, maxIter = 15)
opts_po_cont = ContinuationPar(dsmin = 0.001, dsmax = 0.04, ds = 0.03, pMax = 2.2, maxSteps = 200, newtonOptions = opt_po, saveSolEveryStep = 2,
	plotEveryStep = 1, nev = 11, precisionStability = 1e-6,
	detectBifurcation = 3, dsminBisection = 1e-6, maxBisectionSteps = 15, tolBisectionEigenvalue = 0.)
# number of time slices for the periodic orbit
M = 51
probFD = PeriodicOrbitTrapProblem(M = M)
br_po, = continuation(
	# arguments for branch switching from the first
	# Hopf bifurcation point
	jet..., br, 1,
	# arguments for continuation
	opts_po_cont, probFD;
	# OPTIONAL parameters
	# we want to jump on the new branch at phopf + δp
	# ampfactor is a factor to increase the amplitude of the guess
	δp = 0.01, ampfactor = 1,
	# specific method for solving linear system
	# of Periodic orbits with trapeze method
	# You could use the default one :FullLU (slower here)
	linearPO = :FullSparseInplace,
	# regular options for continuation
	verbosity = 3,	plot = true,
	plotSolution = (x, p; kwargs...) -> heatmap!(reshape(x[1:end-1], 2*n, M)'; ylabel="time", color=:viridis, kwargs...),
	normC = norminf)

Using the above call, it is very easy to find the first branches:

We note that there are several branch points (blue points) on the above diagram. This means that there are additional branches in the neighborhood of these points. We now turn to automatic branch switching on these branches. This functionality, as we shall see, is only provided for PeriodicOrbitTrapProblem.

Let's say we want to branch from the first branch point of the first curve pink branch. The syntax is very similar to the previous one:

br_po2, = BK.continuationPOTrapBPFromPO(
	# arguments for branch switching
	br_po, 1,
	# arguments for continuation
	opts_po_cont; linearPO = :FullSparseInplace,
	ampfactor = 1., δp = 0.01,
	verbosity = 3,	plot = true,
	plotSolution = (x, p; kwargs...) -> heatmap!(reshape(x[1:end-1], 2*n, M)'; ylabel="time", color=:viridis, kwargs...),
	normC = norminf)

It is now straightforward to get the full following diagram

Computation of the branch of periodic orbits (Standard Shooting)

Note that what follows is not really optimized on the DifferentialEquations.jl side. Indeed, we do not use automatic differentiation, we do not pass the sparsity pattern, ...

We now turn to a different method based on the flow of the Brusselator. To compute this flow (time stepper), we need to be able to solve the differential equation (actually a PDE) associated to the vector field Fbru. We will show how to do this with an implicit method Rodas4P from DifferentialEquations.jl. Note that the user can pass its own time stepper but for convenience, we use the ones in DifferentialEquations.jl. More information regarding the shooting method is contained in Periodic orbits based on the shooting method. To define the flow, it is better to have an inplace version of the vector field:

function Fbru!(f, x, p)
	@unpack α, β, D1, D2, l = p
	n = div(length(x), 2)
	h = 1.0 / n; h2 = h*h
	c1 = D1 / l^2 / h2
	c2 = D2 / l^2 / h2

	u = @view x[1:n]
	v = @view x[n+1:2n]

	# Dirichlet boundary conditions
	f[1]   = c1 * (α	  - 2u[1] + u[2] ) + α - (β + 1) * u[1] + f1(u[1], v[1])
	f[end] = c2 * (v[n-1] - 2v[n] + β / α)			 + β * u[n] - f1(u[n], v[n])

	f[n]   = c1 * (u[n-1] - 2u[n] +  α   ) + α - (β + 1) * u[n] + f1(u[n], v[n])
	f[n+1] = c2 * (β / α  - 2v[1] + v[2])			 + β * u[1] - f1(u[1], v[1])

	for i=2:n-1
		  f[i] = c1 * (u[i-1] - 2u[i] + u[i+1]) + α - (β + 1) * u[i] + f1(u[i], v[i])
		f[n+i] = c2 * (v[i-1] - 2v[i] + v[i+1])			  + β * u[i] - f1(u[i], v[i])
	end
	return f
end

Fbru(x, p) = Fbru!(similar(x), x, p)

We then recompute the locus of the Hopf bifurcation points using the same method as above.

n = 100

# different parameters to define the Brusselator model and guess for the stationary solution
par_bru = (α = 2., β = 5.45, D1 = 0.008, D2 = 0.004, l = 0.3)
sol0 = vcat(par_bru.α * ones(n), par_bru.β/par_bru.α * ones(n))

eigls = EigArpack(1.1, :LM)
opts_br_eq = ContinuationPar(dsmin = 0.001, dsmax = 0.00615, ds = 0.0061, pMax = 1.9,
	detectBifurcation = 3, nev = 21, plotEveryStep = 50,
	newtonOptions = NewtonPar(eigsolver = eigls, tol = 1e-9), maxSteps = 200)

br, = @time continuation(Fbru, Jbru_sp,
	sol0, par_bru, (@lens _.l), opts_br_eq, verbosity = 0,
	plot = false,
	printSolution = (x, p)->x[n÷2], normC = norminf)

We need to build a problem which encodes the Shooting functional. This done as follows where we first create the time stepper:

using DifferentialEquations, DiffEqOperators

FOde(f, x, p, t) = Fbru!(f, x, p)

u0 = sol0 .+ 0.01 .* rand(2n)

# this is the ODE time stepper when used with `solve`
probsundials = ODEProblem(FOde, u0, (0., 1000.), par_bru;
	atol = 1e-10, rtol = 1e-8, jac = (J,u,p,t) -> J .= Jbru_sp(u,p), jac_prototype = Jbru_sp(u0, par_bru))
Performance

You can really speed this up by using the improved ODEProblem

using SparseDiffTools, SparseArrays, DiffEqDiffTools
jac_prototype = Jbru_sp(ones(2n), par_bru)
jac_prototype.nzval .= ones(length(jac_prototype.nzval))
_colors = matrix_colors(jac_prototype)
vf = ODEFunction(FOde; jac_prototype = jac_prototype, colorvec = _colors)
probsundials = ODEProblem(vf,  u0, (0.0, 520.), par_bru) # gives 0.22s

We also compute with automatic differentiation, the differentials of the vector field. This is is needed for branch switching as it is based on the computation of the Hopf normal form:

using ForwardDiff
function D(f, x, p, dx)
	return ForwardDiff.derivative(t->f(x .+ t .* dx, p), 0.)
end
d1Fbru(x,p,dx1) = D((z, p0) -> Fbru(z, p0), x, p, dx1)
d2Fbru(x,p,dx1,dx2) = D((z, p0) -> d1Fbru(z, p0, dx1), x, p, dx2)
d3Fbru(x,p,dx1,dx2,dx3) = D((z, p0) -> d2Fbru(z, p0, dx1, dx2), x, p, dx3)

jet  = (Fbru, Jbru_sp, d2Fbru, d3Fbru)

We are now ready to call the automatic branch switching. Note how similar it is to the previous section based on finite differences. This case is more deeply studied in the tutorial Brusselator 1d (advanced user). We use a parallel Shooting.

# linear solvers
ls = GMRESIterativeSolvers(reltol = 1e-7, maxiter = 100, verbose = false)
eig = EigKrylovKit(tol= 1e-12, x₀ = rand(2n), verbose = 0, dim = 40)
# newton parameters
optn_po = NewtonPar(verbose = true, tol = 1e-7,  maxIter = 25, linsolver = ls, eigsolver = eig)
# continuation parameters
opts_po_cont = ContinuationPar(dsmax = 0.03, ds= 0.01, pMax = 2.5, maxSteps = 10,
	newtonOptions = optn_po, nev = 15, precisionStability = 1e-3,
	detectBifurcation = 0, plotEveryStep = 2)

Mt = 2 # number of shooting sections
br_po, = continuation(
	jet..., br, 1,
	# arguments for continuation
	opts_po_cont,
	# this is where we tell that we want Parallel Standard Shooting
	ShootingProblem(Mt, par_bru, probsundials, Rodas4P(), abstol = 1e-10, retol = 1e-8, parallel = true);
	ampfactor = 1.0, δp = 0.0075,
	# the next option is not necessary
	# it speeds up the newton iterations
	# by combining the linear solves of the bordered linear system
	linearAlgo = MatrixFreeBLS(@set ls.N = 2+2n*Mt),
	verbosity = 3,	plot = true,
	plotSolution = (x, p; kwargs...) -> BK.plotPeriodicShooting!(x[1:end-1], Mt; kwargs...),
	normC = norminf)

and you should see

Computation of the branch of periodic orbits (Poincaré Shooting)

We now turn to another Shooting method, namely the Poincaré one. We can provide this method thanks to the unique functionalities of DifferentialEquations.jl. More information is provided at PoincareShootingProblem and Periodic orbits based on the shooting method but basically, it is a shooting method between Poincaré sections $\Sigma_i$ (along the orbit) defined by hyperplanes. As a consequence, the dimension of the unknowns is $M_{sh}\cdot(N-1)$ where $N$ is the dimension of the phase space. Indeed, each time slice lives in an hyperplane $\Sigma_i$. Additionally, the period $T$ is not an unknown of the method but rather a by-product. However, the method requires the time stepper to find when the flow hits an hyperplane $\Sigma_i$, something called event detection.

We show how to use this method, the code is very similar to the case of the Parallel Standard Shooting:

# linear solvers
ls = GMRESIterativeSolvers(reltol = 1e-8, maxiter = 100)
eig = EigKrylovKit(tol= 1e-12, x₀ = rand(2n-1), verbose = 0, dim = 50)
# newton parameters
optn_po = NewtonPar(verbose = true, tol = 1e-7,  maxIter = 25, linsolver = ls, eigsolver = eig)
# continuation parameters
opts_po_cont = ContinuationPar(dsmax = 0.03, ds= 0.005, pMax = 2.5, maxSteps = 100, newtonOptions = optn_po, nev = 10, precisionStability = 1e-5, detectBifurcation = 0, plotEveryStep = 2)

# number of time slices
Mt = 1
br_po, = continuation(
	jet..., br, 1,
	# arguments for continuation
	opts_po_cont, PoincareShootingProblem(Mt, par_bru, probsundials, Rodas4P(); abstol = 1e-10, retol = 1e-8, parallel = true);
	# the next option is not necessary
	# it speeds up the newton iterations
	# by combining the linear solves of the bordered linear system
	linearAlgo = MatrixFreeBLS(@set ls.N = (2n-1)*Mt+1),
	ampfactor = 1.0, δp = 0.005,
	verbosity = 3,	plot = true,
	updateSectionEveryStep = 1,
	plotSolution = (x, p; kwargs...) -> BK.plotPeriodicShooting!(x[1:end-1], Mt; kwargs...),
	normC = norminf)

and you should see: