Internal docstrings

Abstract type for all variables in Jutul.

A variable is associated with a JutulEntity through the associated_entity function. A variable is local to that entity, and cannot depend on other entities. Variables are used by models to define:

• primary variables: Sometimes referred to as degrees of freedom, primary unknowns or solution variables
• parameters: Static quantities that impact the solution
• secondary variables: Can be computed from a combination of other primary and secondary variables and parameters.

Abstract type for scalar variables (one entry per entity, e.g. pressure or temperature in each cell of a model)

Abstract type for vector variables (more than one entry per entity, for example saturations or displacements)

Number of independent primary variables / degrees of freedom per computational entity.

Number of values held by a primary variable. Normally this is equal to the number of degrees of freedom, but some special primary variables are most conveniently defined by having N values and N-1 independent variables.

Upper (inclusive) limit for variable.

Lower (inclusive) limit for variable.

Define a "typical" numerical value for a variable to scale the linear system entries.

Absolute allowable change for variable during a nonlinear update.

Relative allowable change for variable during a nonlinear update. A variable with value |x| and relative limit 0.2 cannot change more than |x|*0.2.

Return the domain entity the equation is associated with

The entity a variable is associated with, and can hold partial derivatives with respect to.

Designate the function as updating a secondary variable.

A generic evaluator is then defined, together with a function for getting the dependencies of that function upon the state. This is most easily documented with an example. If we define the following function annotated with the macro when updating the array containing the values of MyVarType realized for some model:

@jutul_secondary function some_fn!(target, var::MyVarType, model, a, b, c, ix)
    for i in ix
        target[i] = a[i] + b[i] / c[i]
    end
end

The purpose of the macro is to translate this into two functions. The first defines for the dependencies of the function with respect to the fields of the state (primary variables, secondary variables and parameters):

function get_dependencies(var::MyVarType, model)
   return (:a, :b, :c)
end

The second function defines a generic version that takes in state, and automatically expands the set of dependencies into getfield calls.

function update_secondary_variable!(array_target, var::MyVarType, model, state, ix)
    some_fn!(array_target, var, model, state.a, state.b, state.c, ix)
end

Note that the input names of arguments 4 to end-1 matter, as these will be fetched from state, exactly as written.

Get dependencies of variable when viewed as a secondary variable. Normally autogenerated with @jutul_secondary

Super-type for all entities where JutulVariables can be defined.

Entity for Cells (closed volumes with averaged properties for a finite-volume solver)

Entity for Faces (intersection between pairs of Cells)

Entity for Nodes (intersection between multiple Faces)

number_of_partials_per_entity(model::SimulationModel, entity::JutulEntity)

Get the number of local partial derivatives per entity in a model for a given JutulEntity. This is the sum of degrees_of_freedom_per_entity for all primary variables defined on entity.

Get the number of entities (e.g. the number of cells) that the equation is defined on.

Get number of entities a cache is defined on.

Number of entities for vector stored in state (just the number of elements)

Number of entities for matrix stored in state (convention is number of columns)

Number of entities (e.g. Cells, Faces) a variable is defined on. By default, each primary variable exists on all cells of a discretized domain

Abstract type for all residual equations

Get the total number of equations on the domain of model.

Get the number of equations per entity. For example, mass balance of two components will have two equations per grid cell (= entity)

Take value of AD.

value(d::Dict)

Call value on all elements of some Dict.

Create a mapped array that produces only the values when indexed.

Only useful for AD arrays, otherwise it does nothing.

local_ad(state::T, index::I, ad_tag::∂T) where {T, I<:Integer, ∂T}

Create localad for state for index I of AD tag of type adtag

An AutoDiffCache is a type that holds both a set of AD values and a map into some global Jacobian.

Cache that holds an AD vector/matrix together with their positions.

allocate_array_ad(n[, m]; <keyword arguments>)

Allocate vector or matrix as AD with optionally provided context and a specified non-zero on the diagonal.

Arguments

• n::Integer: number of entries in vector, or number of rows if m is given.
• m::Integer: number of rows (optional)

Keyword arguments

• npartials = 1: Number of partials derivatives to allocate for each element
• diag_pos = nothing: Indices of where to put entities on the diagonal (if any)

Other keyword arguments are passed onto get_ad_entity_scalar.

Examples:

Allocate a vector with a single partial:

julia> allocate_array_ad(2)
2-element Vector{ForwardDiff.Dual{nothing, Float64, 1}}:
 Dual{nothing}(0.0,0.0)
 Dual{nothing}(0.0,0.0)

Allocate a vector with two partials, and set the first to one:

julia> allocate_array_ad(2, diag_pos = 1, npartials = 2)
2-element Vector{ForwardDiff.Dual{nothing, Float64, 2}}:
 Dual{nothing}(0.0,1.0,0.0)
 Dual{nothing}(0.0,1.0,0.0)

Set up a matrix with two partials, where the first column has partials [1, 0] and the second [0, 1]:

julia> allocate_array_ad(2, 2, diag_pos = [1, 2], npartials = 2)
2×2 Matrix{ForwardDiff.Dual{nothing, Float64, 2}}:
 Dual{nothing}(0.0,1.0,0.0)  Dual{nothing}(0.0,1.0,0.0)
 Dual{nothing}(0.0,0.0,1.0)  Dual{nothing}(0.0,0.0,1.0)

allocate_array_ad(v::AbstractVector, ...)

Convert vector to AD vector.

allocate_array_ad(v::AbstractMatrix, ...)

Convert matrix to AD matrix.

get_ad_entity_scalar(v::Real, npartials, diag_pos = nothing; <keyword_arguments>)

Get scalar with partial derivatives as AD instance.

Arguments

• v::Real: Value of AD variable.
• npartials: Number of partial derivatives each AD instance holds.
• diag_pos = nothing: Position(s) of where to set 1 as the partial derivative instead of zero.

Keyword arguments

• tag = nothing: Tag for AD instance. Two AD values of the different tag cannot interoperate to avoid perturbation confusion (see ForwardDiff documentation).

Get the entries of the main autodiff cache for an equation.

Note: This only gets the .equation field's entries.

Get entries of autodiff cache. Entries are AD vectors that hold values and derivatives.

Abstract type for matrix layouts. A layout determines how primary variables and equations are ordered in a sparse matrix representation. Note that this is different from the matrix format itself as it concerns the ordering itself: For example, if all equations for a single cell come in sequence, or if a single equation is given for all entities before the next equation is written.

Different layouts does not change the solution of the system, but different linear solvers support different layouts.

Equations are grouped by entity, listing all equations and derivatives for entity 1 before proceeding to entity 2 etc.

For a test system with primary variables P, S and equations E1, E2 and two cells this will give the following ordering on the diagonal: (∂E1/∂p)₁, (∂E2/∂S)₁, (∂E1/∂p)₂, (∂E2/∂S)₂

Equations are stored sequentially in rows, derivatives of same type in columns:

For a test system with primary variables P, S and equations E1, E2 and two cells this will give the following ordering on the diagonal: (∂E1/∂p)₁, (∂E1/∂p)₂, (∂E2/∂S)₁, (∂E2/∂S)₂

Same as EntityMajorLayout, but the system is a sparse matrix where each entry is a small dense matrix.

For a test system with primary variables P, S and equations E1, E2 and two cells this will give a diagonal of length 2: [(∂E1/∂p)₁ (∂E1/∂S)₁ ; (∂E2/∂p)₁ (∂E2/∂S)₁] [(∂E1/∂p)₂ (∂E1/∂S)₂ ; (∂E2/∂p)₂ (∂E2/∂S)₂]

convergence_criterion(model, storage, eq, eq_s, r; dt = 1)

Get the convergence criterion values for a given equation. Can be checked against the corresponding tolerances.

Arguments

• model: model that generated the current equation.
• storage: global simulator storage.
• eq::JutulEquation: equation implementation currently being checked
• eq_s: storage for eq where values are contained.
• r: the local residual part corresponding to this model, as a matrix with column index equaling entity index

partition(N::AbstractMatrix, num_coarse, weights = ones(size(N, 2)); partitioner = MetisPartitioner(), groups = nothing, n = maximum(N), group_by_weights = false, buffer_group = true)

Partition based on neighborship (with optional groups kept contigious after partitioning)

load_balanced_endpoint(block_index, nvals, nblocks)

Endpoint for interval block_index that subdivides nvals into nblocks in a load balanced manner. This is done by adding one element to the first set of blocks whenever possible.