Time-stepping
For convenience, we include utilities for two explicit Runge-Kutta schemes. For more advanced time-stepping functionality, we recommend DifferentialEquations.jl
.
Carpenter and Kennedy's (4,5) method
ck45()
returns coefficients for a low-storage Runge-Kutta method.
Example usage:
using Plots
using StartUpDG
# Brusselator
B = 3
f(y1,y2) = [1+y1^2*y2-(B+1)*y1, B*y1-y1^2*y2]
f(Q) = f(Q[1],Q[2])
p,q = 1.5, 3.0
Q = [p;q]
dt = .1
FinalTime = 20
res = zero.(Q) # init RK residual
rk4a,rk4b,rk4c = ck45()
Nsteps = ceil(FinalTime/dt)
dt = FinalTime/Nsteps
plot() # init plot
for i = 1:Nsteps
global res # yes, I know...this is just for simplicty
for INTRK=1:5
rhsQ = f(Q)
@. res = rk4a[INTRK]*res + dt*rhsQ # i = RK stage
@. Q = Q + rk4b[INTRK]*res
end
scatter!([i*dt;i*dt],Q)
end
display(plot!(leg=false))
Dormand-Prince (5,6) with PI error control
dp56()
returns coefficients for an embedded Runge-Kutta method.
Example usage:
using Plots
using StartUpDG
# Brusselator
B = 3
f(y1,y2) = [1+y1^2*y2-(B+1)*y1, B*y1-y1^2*y2]
f(Q) = f(Q[1],Q[2])
p,q = 1.5, 3.0
Q = [p;q]
dt = .01
FinalTime = 20
rka,rkE,rkc = dp56()
PI = PIparams(order=5,errTol=1e-6)
Qtmp = copy(Q) # temp storage for RK stage vecs y_i
rhsQrk = (f(Q),ntuple(x->zero(Q),length(rkE)-1)...) # RK f(y_i), set first elem (FSAL)
prevErrEst = nothing
t = 0.0
plot()
while t < FinalTime
global Q,rhsQrk,t,dt,prevErrEst
for INTRK = 2:7
fill!(Qtmp,zero(eltype(first(Q))))
for s = 1:INTRK-1
@. Qtmp = Qtmp + rka[INTRK,s]*rhsQrk[s]
end
@. Qtmp = Q + dt*Qtmp
rhsQrk[INTRK] .= f(Qtmp)
end
# Hairer seminorm error estimator
errEstVec = sum(rkE.*rhsQrk) ./ (PI.errTol * (@. 1 + abs(Q)))
errEst = mapreduce(x->x*x,+,errEstVec) / sum(length.(Q))
accept_step,dtnew,prevErrEst = compute_adaptive_dt(errEst,dt,PI,prevErrEst)
if accept_step
t += dt
@. Q = Qtmp
@. rhsQrk[1] = rhsQrk[7] # use FSAL property
scatter!([t;t],Q)
end
dt = min(FinalTime-t,dtnew)
end
display(plot!(leg=false))