When Your Modal Model Won't Fit: Craig-Bampton Substructuring in Ansys Mechanical
Craig-Bampton substructuring in Ansys Mechanical condenses large assemblies into reduced modal models. Here's the workflow, the accuracy trade-offs, and how to verify results against a full solve.
23 Mar 2026, 00:34 UTC

A full-vehicle modal model can easily run past ten million degrees of freedom. You mesh it, wait hours for the eigenvalue solve, get your first twenty modes — and then the design team changes a bracket and you start over. If that loop sounds familiar, Craig-Bampton substructuring in Ansys Mechanical is worth your time: it lets you condense each assembly section into a small reduced model, assemble the reduced pieces, and still recover natural frequencies within a few percent of the full solve, often with the majority of the solve time gone.
What Craig-Bampton actually does
Component Mode Synthesis (CMS) is the family name; Craig-Bampton is the specific variant Ansys implements. The idea is to split a large structure into substructures (also called superelements) and represent each one with two kinds of modes:
- Fixed-interface modes — the natural modes of the substructure with its boundary degrees of freedom (DOFs) clamped. These capture interior flexibility.
- Constraint modes — static shapes produced by giving each interface DOF a unit displacement while holding the others fixed. These guarantee the substructures can still connect and transmit loads correctly at their boundaries.
The substructure's mass and stiffness matrices are projected onto this small modal basis, so a body section with two million DOFs might collapse to a few hundred. The reduced matrices are then assembled at the shared interfaces and the modal solve runs on the condensed system. The physics you keep is exact at the interfaces and approximate in the interior — which is precisely where the trade-off lives.
Workflow in Mechanical (2021 R2 and later)
The automated workflow arrived in Ansys Mechanical 2021 R2; on earlier releases you would need to drive the reduction with manual APDL command snippets, which is workable but error-prone. Assuming a current release, the workflow in Workbench looks like this:
- Open a Modal analysis and insert a Condensed Geometry object for each substructure you want to reduce (right-click Model → Insert → Condensed Geometry, then scope the bodies).
- On each condensed part, define the interface — the faces or remote points where it connects to neighboring substructures.
- Set the number of fixed-interface modes to retain. A starting range of 10–20 per substructure is common; the right number depends on the frequency band you care about (more on that below).
- Solve. Mechanical generates the reduced matrices, assembles them, and runs the modal solve on the condensed system in one pass.
A practical bonus: the reduced matrices for a substructure can be exported and reused. If the suspension substructure doesn't change across five body variants, you reduce it once and reuse it in each variant's solve — this is where the parametric-study payoff really compounds.
A worked example: passenger-car chassis
Consider a chassis model split into three substructures: body-in-white, suspension cradle, and powertrain. The full model solves as one monolithic eigenproblem. The substructured version retains, say, 15 fixed-interface modes per substructure plus the constraint modes at each interface.
Typical results reported for this class of model: the first six global natural frequencies land within roughly 3% of the full-model solution, while solve time drops on the order of 70%. Those numbers are illustrative of what the method can deliver, not a guarantee — your accuracy depends entirely on mode selection and how much local deformation matters in your frequency band. Treat them as a reason to try the method, then verify on your own geometry.
The trade-off: how many modes is enough?
Retained modes are your accuracy budget. Too few fixed-interface modes and you miss local flexibility — a thin panel that breathes at 180 Hz in the full model may simply not exist in the reduced basis, and nearby global modes shift or vanish. Too many, and you erode the time savings.
A defensible selection heuristic:
- Identify the highest frequency you need to capture accurately, fmax.
- Retain fixed-interface modes up to roughly 1.5–2× fmax per substructure. The margin absorbs truncation error.
- If a substructure has known local modes in your band (brackets, panels, shafts), bias the count upward for that substructure specifically.
Also be clear about scope: Craig-Bampton is a linear reduction. If your assembly's behavior depends on contact status, bolted-joint slip, or material nonlinearity, the reduced model inherits whatever linearization you imposed — or misrepresents the physics entirely. Nonlinear or contact-dominated problems generally need a full model or a different reduction strategy.
Verification you should actually run
Never trust a substructured result you haven't benchmarked, at least once per model family:
- Run the full-model modal analysis with identical boundary conditions and extract the first N frequencies and mode shapes.
- Run the substructured model and compare. A common acceptance target is frequency differences under 5% and Modal Assurance Criterion (MAC) values above 0.9 for paired mode shapes. MAC is a 0–1 correlation measure between two mode-shape vectors; below ~0.9, you're likely looking at a different or distorted mode.
- Check interface continuity: inspect displacements and reaction forces at the substructure boundaries. Discontinuities there indicate an interface definition problem, not a mode-count problem.
If frequencies match but MAC is poor for a specific mode, that mode probably involves local deformation your basis truncated — add fixed-interface modes to the offending substructure and re-solve.
Closing: when to reach for it
Use Craig-Bampton when your model is large, linear, and iterated on — vehicle bodies, airframe sections, rotating machinery housings — and especially when only part of the assembly changes between design loops. Skip it when the problem is small enough to solve directly, or when nonlinear contact drives the answer. The actionable first step: take a model you already trust, rebuild it with one condensed substructure, and run the frequency/MAC comparison above. That single benchmark tells you more about whether the method fits your workflow than any feature description will.
0 replies
A thoughtful contribution can make all the difference. Be the first to share one.