4. Nonlinear solution strategies

In nonlinear finite element analysis, the equilibrium equations are no longer linear in the displacement vector. For a discretized structure, the residual force vector at iteration i is defined as R(u_i) = F_ext − F_int(u_i), where F_int is assembled from element internal forces that depend nonlinearly on the current displacement state. The goal of every solution strategy in this chapter is to drive R to zero along a prescribed load path, while remaining stable when that path becomes complex—snap-through, snap-back, material softening, contact activation, or local buckling.

The choice of strategy is not cosmetic. A poorly chosen scheme can converge to the wrong equilibrium branch, stall at a limit point, or waste hours of CPU time oscillating between two states. This section compares the four workhorse methods—full Newton–Raphson, modified Newton, arc-length, and line search—and shows how to combine them in practice.

Newton–Raphson: The Reference Method

The full Newton–Raphson (NR) method linearizes the residual at each iteration and solves for an incremental displacement correction:

K_T(u_i) · Δu = R(u_i)
u_(i+1) = u_i + Δu

Here K_T is the tangent stiffness matrix, reassembled and refactorized at every iteration. The quadratic convergence near the solution is the payoff: once the iterate enters the convergence radius, the residual typically drops by several orders of magnitude in two or three iterations. The cost is a full assembly and factorization per iteration, which dominates runtime for large models.

For complex load paths, full NR is the most reliable choice when the tangent is well-conditioned and the load steps are small enough to stay on the same equilibrium branch. It is the default in most implicit solvers (Abaqus/Standard, ANSYS Mechanical, Nastran SOL 400) and should be your baseline before considering alternatives.

Modified Newton: Trading Convergence Rate for Cost

Modified Newton (mNR) reuses the tangent stiffness from the beginning of the step (or from a chosen iteration) and only updates the residual:

K_T(u_0) · Δu_i = R(u_i)
u_(i+1) = u_i + Δu_i

This eliminates repeated factorization, which can cut per-iteration cost by 5–20× on large models. The trade-off is linear (or slower) convergence, so more iterations are needed. mNR is attractive for problems with mild nonlinearity—large-displacement small-strain, hyperelasticity away from instabilities—or when the tangent changes slowly. It is a poor choice near limit points, where the initial tangent becomes nearly singular and convergence stalls.

Arc-Length Methods: Traversing Limit Points

When the load–displacement curve has a limit point (snap-through) or a turning point (snap-back), load-controlled NR fails because no equilibrium exists at the prescribed load level. Arc-length methods (Riks, Crisfield, modified Riks) augment the system with a constraint on the solution increment:

R(u, λ) = F_ext(λ) − F_int(u) = 0
ΔuᵀΔu + Δλ²ψ² = Δs²

Here λ is the load proportionality factor, Δs is the arc-length radius, and ψ is a scaling factor that balances displacement and load contributions. The augmented system is solved for both Δu and Δλ, allowing the solution to pass through limit points where K_T is singular.

Key practical settings:

• Initial arc-length Δs₀: typically 5–10% of a characteristic displacement (e.g., span/100 for a shell).
• Adaptive stepping: if convergence takes < 4 iterations, increase Δs by 1.5×; if > 10 iterations or divergence, reduce by 0.5×.
• Minimum/maximum bounds: set Δs_min = 0.01·Δs₀ and Δs_max = 5·Δs₀ to avoid stalling or overshooting.
• Sign of Δλ: determined by the sign of the tangent stiffness determinant; a negative determinant indicates traversal past a limit point.

Arc-length is essential for post-buckling analysis, snap-through of shallow arches, and softening responses in concrete or composites. It is not a cure-all: for path-dependent problems with sharp local instabilities (e.g., contact chatter), the arc-length constraint can produce non-physical load reversals. In such cases, combine arc-length with line search or switch to a dynamic relaxation scheme.

Line Search: Stabilizing the Iteration

Even with a good tangent, the full NR update Δu may overshoot the equilibrium basin, causing divergence or oscillation. A line search scales the update by a factor η ∈ (0, 1] to minimize the residual norm along the search direction:

u_(i+1) = u_i + η · Δu
minimize ||R(u_i + η · Δu)||²

The most common implementation is the Armijo backtracking line search: start with η = 1, and while ||R(u_i + ηΔu)|| > (1 − cη)||R(u_i)|| (with c ≈ 1e-4), set η ← η/2. This guarantees a sufficient decrease in the residual norm and is cheap—only residual evaluations, no additional factorizations.

Line search is particularly effective for:

• Problems with strong material nonlinearity (plasticity, damage) where the tangent is inconsistent.
• Contact problems where the active set changes between iterations.
• Large load steps where the initial guess is far from equilibrium.

Choosing and Combining Strategies

The table below summarizes when each method is appropriate. In practice, modern solvers combine them: full NR with line search for general nonlinearity, arc-length with line search for limit-point problems, and automatic switching between full and modified Newton based on convergence rate.

Worked Example: Snap-Through of a Shallow Arch

Consider a shallow circular arch with span 1000 mm, rise 50 mm, rectangular cross-section 20 mm × 5 mm, linear elastic steel (E = 210 GPa, ν = 0.3). A concentrated load is applied at the crown. The expected response is a snap-through at a limit load of approximately 120 N, followed by a stable post-snap branch.

Step 1—Model setup: Use 2D beam elements (B21 in Abaqus) with 50 elements along the arch. Enable NLGEOM for large displacements. Apply the load as a concentrated force at the crown node.

Step 2—Solution controls: Use the Riks arc-length method with an initial arc-length of 0.5 mm, minimum 0.01 mm, maximum 5 mm. Set a maximum of 100 increments and a convergence tolerance of 1e-5 on the residual.

Step 3—Expected output: The load–displacement curve should show a peak at ~120 N, a sharp drop to ~40 N, and a stable branch thereafter. The arc-length method will automatically reduce the step size near the limit point and continue past it.

Step 4—Verification: Check that the determinant of the tangent stiffness changes sign at the limit point. If the analysis terminates prematurely, reduce the initial arc-length or increase the maximum number of increments. If the solution jumps to a non-physical branch, enable line search or reduce the arc-length further.

Practical Guidelines

• Always start with full NR and a small load step. If it converges in 3–5 iterations, increase the step size gradually.
• Use arc-length when you expect a limit point. Do not wait for NR to fail—switch proactively if the structure is prone to buckling or snap-through.
• Enable line search for contact and plasticity. It is cheap and often prevents divergence without reducing the step size.
• Monitor the convergence history. A sudden increase in iterations or a drop in η indicates a change in the active set or a bad tangent.
• Combine methods. Full NR + line search for general problems; arc-length + line search for limit points; modified Newton only when the tangent is expensive and the nonlinearity is mild.

The figure above contrasts the three solution paths on a snap-through problem. Load-controlled NR follows the equilibrium curve until the limit point, where the tangent stiffness becomes singular and the iteration diverges. Arc-length continues along the curve, automatically reducing the step size near the limit point and traversing the unstable branch. Line search, shown as discrete steps along the NR path, stabilizes the iteration by scaling the update when the full step would overshoot.

In production analyses, the most robust approach is to combine arc-length with line search and full NR. This configuration handles the widest range of complex load paths—from smooth nonlinearity to sharp instabilities—while maintaining quadratic convergence where possible. Always verify the solution by checking the determinant sign, the convergence history, and the physical plausibility of the post-limit branch.