DAE Initialization

DAE (Differential-Algebraic Equation) problems often require special initialization procedures to ensure that the initial conditions are consistent with the algebraic constraints. The DifferentialEquations.jl ecosystem provides several initialization algorithms to handle this automatically or to verify that your provided initial conditions are already consistent.

The Initialization Problem

DAEs have the general form:

\[M \frac{du}{dt} = f(u, p, t)\]

where M is a (possibly singular) mass matrix. For the initial conditions u₀ and du₀ to be consistent, they must satisfy:

\[f(du₀, u₀, p, t₀) = 0\]

for fully implicit DAEs, or the equivalent constraint for semi-explicit DAEs. Finding consistent initial conditions is a nonlinear problem that must be solved before time integration can begin.

Available Initialization Algorithms

The initializealg keyword argument to solve controls how initialization is performed. All algorithms are documented with their docstrings:

struct CheckInit <: DAEInitializationAlgorithm

struct NoInit <: DAEInitializationAlgorithm

struct OverrideInit <: DAEInitializationAlgorithm

⚠️ WARNING: NoInit Usage

Algorithm Selection Guide

Examples

Example 1: Simple Pendulum DAE

using OrdinaryDiffEqBDF: DFBDF
using SciMLBase: DAEProblem, solve
using DiffEqBase: BrownFullBasicInit, CheckInit, NoInit

function pendulum!(res, du, u, p, t)
    x, y, T = u
    dx, dy, dT = du
    g, L = p

    res[1] = dx - du[1]
    res[2] = dy - du[2]
    res[3] = x^2 + y^2 - L^2  # Algebraic constraint
    return
end

u0 = [1.0, 0.0, 0.0]  # Initial position
du0 = [0.0, 0.0, 0.0]  # Initial velocity (inconsistent!)
p = [9.81, 1.0]  # g, L
tspan = (0.0, 10.0)

prob = DAEProblem(
    pendulum!, du0, u0, tspan, p,
    differential_vars = [true, true, false]
)

# BrownFullBasicInit will fix the inconsistent du0
sol = solve(prob, DFBDF(), initializealg = BrownFullBasicInit())

Example 2: Checking Consistency (Recommended over NoInit)

# If you've computed consistent conditions yourself
u0_consistent = [1.0, 0.0, 0.0]
du0_consistent = [0.0, -1.0, compute_tension(u0_consistent, p)]

prob2 = DAEProblem(
    pendulum!, du0_consistent, u0_consistent, tspan, p,
    differential_vars = [true, true, false]
)

# RECOMMENDED: Verify they're consistent with CheckInit
sol = solve(prob2, DFBDF(), initializealg = CheckInit())

# NOT RECOMMENDED: NoInit skips all checks - use at your own risk!
# sol = solve(prob2, DFBDF(), initializealg = NoInit())  # ⚠️ DANGEROUS

Example 3: ModelingToolkit Integration

When using ModelingToolkit, initialization information is often included automatically:

using ModelingToolkit
using OrdinaryDiffEqBDF: DFBDF
using SciMLBase: DAEProblem, solve

@variables t x(t) y(t) T(t)
@parameters g L
D = Differential(t)

eqs = [
    D(x) ~ -T * x / L,
    D(y) ~ -T * y / L - g,
    x^2 + y^2 ~ L^2,
]

@named pendulum = ODESystem(eqs, t, [x, y, T], [g, L])
sys = structural_simplify(pendulum)

# ModelingToolkit provides initialization_data
prob = DAEProblem(sys, [x => 1.0, y => 0.0], (0.0, 10.0), [g => 9.81, L => 1.0])

# DefaultInit will use OverrideInit with ModelingToolkit's initialization_data
sol = solve(prob, DFBDF())  # Automatic initialization!

Algorithm-Specific Options

Both OrdinaryDiffEq and Sundials support the same initialization algorithms through the initializealg parameter:

OrdinaryDiffEq and Sundials

using OrdinaryDiffEqBDF: DFBDF
# or
using Sundials: IDA

# Use Brown's algorithm to fix inconsistent conditions
sol = solve(prob, DFBDF(), initializealg = BrownFullBasicInit())  # OrdinaryDiffEq
sol = solve(prob, IDA(), initializealg = BrownFullBasicInit())    # Sundials

# Use Shampine's collocation method for general DAEs
sol = solve(prob, DFBDF(), initializealg = ShampineCollocationInit())  # OrdinaryDiffEq
sol = solve(prob, IDA(), initializealg = ShampineCollocationInit())    # Sundials

# RECOMMENDED: Verify conditions are consistent
sol = solve(prob, DFBDF(), initializealg = CheckInit())  # OrdinaryDiffEq
sol = solve(prob, IDA(), initializealg = CheckInit())    # Sundials

# NOT RECOMMENDED: Skip all initialization checks (dangerous!)
# sol = solve(prob, DFBDF(), initializealg = NoInit())  # ⚠️ USE AT YOUR OWN RISK
# sol = solve(prob, IDA(), initializealg = NoInit())    # ⚠️ USE AT YOUR OWN RISK

Troubleshooting

Common Issues and Solutions

1. "Initial conditions are not consistent" error

   • Ensure your du0 satisfies the DAE constraints at t0
   • Try using BrownFullBasicInit() or ShampineCollocationInit() instead of CheckInit()
   • Check that differential_vars correctly identifies differential vs algebraic variables

2. Initialization fails to converge

   • Relax tolerances if using extended versions
   • Try a different initialization algorithm
   • Provide a better initial guess for algebraic variables
   • Check if your DAE is index-1: The system may be higher-index (see below)

3. Solver fails immediately after initialization

   • The initialization might have found a consistent but numerically unstable point
   • Try tightening initialization tolerances
   • Check problem scaling and consider non-dimensionalization

4. DAE is not index-1 (higher-index DAE)

   • Many initialization algorithms only work reliably for index-1 DAEs
   • To check if your DAE is index-1: The Jacobian of the algebraic equations with respect to the algebraic variables must be non-singular
   • Solution: Use ModelingToolkit.jl to analyze and potentially reduce the index:

   using ModelingToolkit
   
   # Define your system with ModelingToolkit
   @named sys = ODESystem(eqs, t, vars, params)
   
   # Analyze and reduce the index (structural_simplify handles this in v10+)
   sys_reduced = structural_simplify(sys)
   
   # The reduced system will be index-1 and easier to initialize
   prob = DAEProblem(sys_reduced, [], (0.0, 10.0), params)

   • ModelingToolkit can automatically detect the index and apply appropriate transformations
   • After index reduction, standard initialization algorithms will work more reliably

Performance Tips

1. Use differential_vars when possible: This helps initialization algorithms understand problem structure
2. Provide good initial guesses: Even when using automatic initialization, starting closer to the solution helps
3. Consider problem-specific initialization: For complex systems, custom initialization procedures may be more efficient
4. Use CheckInit() when appropriate: If you know conditions are consistent, skip unnecessary computation

References

• Brown, P. N., Hindmarsh, A. C., & Petzold, L. R. (1998). Consistent initial condition calculation for differential-algebraic systems. SIAM Journal on Scientific Computing, 19(5), 1495-1512.
• Shampine, L. F. (2002). Consistent initial condition for differential-algebraic systems. SIAM Journal on Scientific Computing, 22(6), 2007-2026.