Merge pull request #861 from SciML/sccnonlinear1

Add SCCNonlinearProblem

src/SciMLBase.jl

@@ -791,9 +791,10 @@ export isinplace
 
 export solve, solve!, init, discretize, symbolic_discretize
 
-export LinearProblem, NonlinearProblem, IntervalNonlinearProblem,
-       IntegralProblem, SampledIntegralProblem, OptimizationProblem,
-       NonlinearLeastSquaresProblem
+export LinearProblem, IntervalNonlinearProblem,
+       IntegralProblem, SampledIntegralProblem, OptimizationProblem
+
+export NonlinearProblem, SCCNonlinearProblem, NonlinearLeastSquaresProblem
 
 export DiscreteProblem, ImplicitDiscreteProblem
 export SteadyStateProblem, SteadyStateSolution

src/problems/nonlinear_problems.jl

@@ -321,3 +321,148 @@ end
 function NonlinearLeastSquaresProblem(f, u0, p = NullParameters(); kwargs...)
     return NonlinearLeastSquaresProblem(NonlinearFunction(f), u0, p; kwargs...)
 end
+
+@doc doc"""
+    SCCNonlinearProblem(probs, explicitfuns!)
+
+Defines an SCC-split nonlinear system to be solved iteratively.
+
+## Mathematical Specification of an SCC-Split Nonlinear Problem
+
+An SCC-Split Nonlinear Problem is a system of nonlinear equations
+
+```math
+f(u,p) = 0
+```
+
+with the special property that its Jacobian is in block-lower-triangular
+form. In this form, the nonlinear problem can be decomposed into a system
+of nonlinear systems. 
+
+```math
+f_1(u_1,p) = 0
+f_2(u_2,u_1,p) = 0
+f_3(u_3,u_2,u_1,p) = 0
+\vdots
+f_n(u_n,\ldots,u_3,u_2,u_1,p) = 0
+```
+
+Splitting the system in this form can have multiple advantages, including:
+
+* Improved numerical stability and robustness of the solving process
+* Improved performance due to using smaller Jacobians
+
+The SCC-Split Nonlinear Problem is the ordered collection of nonlinear systems 
+to solve in order solve the system in the optimized split form.
+
+## Representation
+
+The representation of the SCCNonlinearProblem is via an ordered collection
+of `NonlinearProblem`s, `probs`, with an attached explicit function for pre-processing
+a cache. This can be interpreted as follows:
+
+```math
+p_1 = g_1(u,p)
+f_1(u_1,p_1) = 0
+p_2 = g_2(u,p)
+f_2(u_2,u_1,p_2) = 0
+p_3 = g_3(u,p)
+f_3(u_3,u_2,u_1,p_3) = 0
+\vdots
+p_n = g_n(u,p)
+f_n(u_n,\ldots,u_3,u_2,u_1,p_n) = 0
+```
+
+where ``g_i`` is `explicitfuns![i]`. In a computational sense, `explictfuns!`
+is instead a mutating function `explictfuns![i](prob.probs[i],sols)` which
+updates the values of prob.probs[i] using the previous solutions `sols[i-1]`
+and below.
+
+!!! warn
+    For the purposes of differentiation, it's assumed that `explictfuns!` does
+    not modify tunable parameters!
+
+!!! warn
+    While `explictfuns![i]` could in theory use `sols[i+1]` in its computation,
+    these values will not be updated. It is thus the contract of the interface
+    to not use those values except for as caches to be overriden.
+
+!!! note
+    prob.probs[i].p can be aliased with each other as a performance / memory
+    optimization. If they are aliased before the construction of the `probs`
+    the runtime guarantees the aliasing behavior is kept.
+
+## Example
+
+For the following nonlinear problem:
+
+```julia
+function f(du,u,p)
+	du[1] = cos(u[2]) - u[1]
+	du[2] = sin(u[1] + u[2]) + u[2]
+	du[3] = 2u[4] + u[3] + 1.0
+	du[4] = u[5]^2 + u[4]
+	du[5] = u[3]^2 + u[5]
+	du[6] = u[1] + u[2] + u[3] + u[4] + u[5]    + 2.0u[6] + 2.5u[7] + 1.5u[8]
+	du[7] = u[1] + u[2] + u[3] + 2.0u[4] + u[5] + 4.0u[6] - 1.5u[7] + 1.5u[8]
+	du[8] = u[1] + 2.0u[2] + 3.0u[3] + 5.0u[4] + 6.0u[5] + u[6] - u[7] - u[8]
+end
+prob = NonlinearProblem(f, zeros(8))
+sol = solve(prob)
+```
+
+The split SCC form is:
+
+```julia
+cache = zeros(3)
+
+function f1(du,u,cache)
+	du[1] = cos(u[2]) - u[1]
+	du[2] = sin(u[1] + u[2]) + u[2]
+end
+explicitfun1(cache,sols) = nothing
+prob1 = NonlinearProblem(NonlinearFunction{true, SciMLBase.NoSpecialize}(f1), zeros(2), cache)
+sol1 = solve(prob1, NewtonRaphson())
+
+function f2(du,u,cache)
+	du[1] = 2u[2] + u[1] + 1.0
+	du[2] = u[3]^2 + u[2]
+	du[3] = u[1]^2 + u[3]
+end
+explicitfun2(cache,sols) = nothing
+prob2 = NonlinearProblem(NonlinearFunction{true, SciMLBase.NoSpecialize}(f2), zeros(3), cache)
+sol2 = solve(prob2, NewtonRaphson())
+
+function f3(du,u,cache)
+	du[1] = cache[1] + 2.0u[1] + 2.5u[2] + 1.5u[3]
+	du[2] = cache[2] + 4.0u[1] - 1.5u[2] + 1.5u[3]
+	du[3] = cache[3] + + u[1] - u[2] - u[3]
+end
+prob3 = NonlinearProblem(NonlinearFunction{true, SciMLBase.NoSpecialize}(f3), zeros(3), cache)
+function explicitfun3(cache,sols)
+	cache[1] = sols[1][1] + sols[1][2] + sols[2][1] + sols[2][2] + sols[2][3]
+	cache[2] = sols[1][1] + sols[1][2] + sols[2][1] + 2.0sols[2][2] + sols[2][3]
+	cache[3] = sols[1][1] + 2.0sols[1][2] + 3.0sols[2][1] + 5.0sols[2][2] + 6.0sols[2][3]
+end
+explicitfun3(cache,[sol1,sol2])
+sol3 = solve(prob3, NewtonRaphson())
+manualscc = [sol1; sol2; sol3]
+
+sccprob = SciMLBase.SCCNonlinearProblem([prob1,prob2,prob3], SciMLBase.Void{Any}.([explicitfun1,explicitfun2,explicitfun3]))
+```
+
+Note that this example aliases the parameters together for a memory-reduced representation.
+
+## Problem Type
+
+### Constructors
+
+### Fields
+
+* `probs`: the collection of problems to solve
+* `explictfuns!`: the explicit functions for mutating the parameter set
+"""
+mutable struct SCCNonlinearProblem{P, E}
+    probs::P
+    explictfuns!::E
+end

src/solutions/nonlinear_solutions.jl

@@ -70,7 +70,7 @@ const SteadyStateSolution = NonlinearSolution
 
 get_p(p::AbstractNonlinearSolution) = p.prob.p
 
-function build_solution(prob::AbstractNonlinearProblem,
+function build_solution(prob::Union{AbstractNonlinearProblem, SCCNonlinearProblem},
         alg, u, resid; calculate_error = true,
         retcode = ReturnCode.Default,
         original = nothing,