Welding distortion simulation for large fabrications, from weld sequence to pre-set

Welding distortion simulation predicts how far a fabricated structure shrinks, bows and twists as its arc welds cool, so weld sequence, fixtures and pre-set are chosen on a screen, not in the bay. We build it for large GMAW, GTAW and SAW structures: a calibrated moving-source weld model, fast assembly models and a surrogate that ranks weld sequences, on the solver you already use.

The first girder of a new design leaves the welding bay bowed, its end plates out of square with the next section. The straightening crew is booked and the jig needs a redesign. The question for the second unit has no answer on any drawing: which seam goes first, from which end, and how far should the parts be pre-set? Full-size trials answer it slowly, one assembly at a time.

What does welding distortion simulation have to predict?

The result that matters is the shape of the finished assembly at its interfaces: the gap at the next joint, the straightness of a girder, the flatness of a panel. Welding shrinks a structure across the joint and along it, and because contraction is not uniform through the thickness the parts also rotate, which is angular distortion.

The cause is a hot, narrow zone held by a cold, stiff surround. TWI notes that in CMn steel the molten weld metal loses about 3% of its volume on solidification and the solidified weld metal and heat-affected zone a further 7% as they cool to room temperature, and that stress above yield leaves permanent plastic deformation. What remains is a band of plastic strain around each seam, and the rest of the structure responds to it elastically.

Acceptance is a tolerance on the drawing. ISO 13920:2023 gives general tolerances for linear and angular dimensions and for shape and position of welded structures in four classes, and the tolerances that apply are those stated on the drawing. The model gives a predicted deviation at each interface to set beside them.

Why is arc weld distortion so hard to calculate?

Three problems share one model. The first is the moving source. Arc power into the work is Q = ηVI, with efficiency η, voltage V and current I, and heat input per unit length is Q/v at travel speed v. It lands over a volume, not a point: the standard description is the double-ellipsoid source of Goldak, Chakravarti and Bibby, in a frame moving with the arc, where x is across the weld, y is depth into the plate and ξ is distance ahead of the arc:

q_f = 6√3 f_f Q / (π√π a b c_f) · exp[−3x²/a² − 3y²/b² − 3ξ²/c_f²] for ξ ≥ 0, and the same with f_r and c_r behind the arc, with f_f + f_r = 2. The lengths a, b, c_f and c_r are not material properties. They are calibrated until the computed fusion line matches a macrograph of the real weld, so each process, joint and parameter set needs its own.

The second is the filler. The model starts with the groove elements inactive and activates them as the source reaches them. The third is material behaviour that depends on temperature and, in steels, on phase: yield strength falls on heating and recovers on cooling, and the transformation of austenite changes volume and strength, which sets the residual stress that remains. In a V joint the second run rotates the plates about the first deposit, as TWI describes.

How are large welded structures usually handled, and where does that break?

Mostly by experience. TWI gives a rule of thumb for shrinkage allowances and calls the amount almost impossible to predict accurately. TWI on pre-setting and restraint says the same of pre-setting: a number of trial welds will be required, so it suits simple assemblies. Restraint is more widely practised, but it adds residual stress that can crack susceptible materials, and the part still moves a little when it leaves the fixture.

Calculation has its limit too. A moving-source thermal elastic-plastic analysis of every pass is physically complete, but Murakami and co-authors in ISIJ International note that it is not always realistic for production sites and design departments, given the calculation time on large structures such as ships and construction machinery. The shortcut is the inherent strain method and its thermal shrinkage variant: apply the strain left by each weld to a model of the whole assembly, at extremely short calculation time. In the thermal shrinkage form that strain is ε = −αΔT inside a zone around the weld. The catch is the input: the same paper states that a shrinkage strain is needed for each welding condition and each joint, and that setting it is arbitrary. An uncalibrated shortcut is a guess that runs fast. Our article on metal additive manufacturing distortion covers the same idea for layered builds.

Sequence is the third limit. TWI recommends welding towards the free end of the joint, with short back-step or skip runs on long welds. An assembly of r seams, each weldable in either direction, has r! × 2^r orders. Eight seams give 10,321,920, the example Wu and Kim use in their review of FEM-based distortion mitigation, where a genetic algorithm is the most common search and thousands of thermo-elastic-plastic cases make computation time the obstacle.

How do you build a weld model your engineers can trust?

We split the problem the way the table does. On a representative joint we build the full model in LS-DYNA, Ansys, Abaqus or your existing thermal-structural solver: source, element activation, temperature- and phase-dependent properties, and the passes as you weld them, lumped into one step only where the validation record shows it holds. The thermal result is verified first, fusion line against the macrograph and heat balance against ηVI/v, then mesh and time-step convergence, then the measured distortion of a test piece. The strain each seam leaves goes into a library that the assembly model reads. Our arc welding and welding simulation pages set the wider context.

Uncertainty is stated, not hidden. Arc efficiency, heat-source lengths, high-temperature material data, fit-up and restraint all move the answer, so we sweep them and report distortion as a range with its assumptions. Residual stress from a critical joint then passes to your fatigue life assessment with its range.

Where does AI help weld sequence optimisation, and where does it not?

• Sequence ranking. A designed set of orders runs through the assembly model, and a surrogate trained on those runs predicts interface distortion for any order. The search over millions of orders runs on the surrogate, and only the best few return to the full model.
• Calibration. Fitting heat-source lengths and shrinkage inputs to the measured fusion line and distortion is an inverse problem that automated fitting makes repeatable.
• Run triage. Checks reject non-converged or energy-unbalanced runs before they reach the training set.

It does not replace the physics. A surrogate predicts only inside the joint types, grades and process window it was trained on, inherits the errors of the model behind it, and cannot tell that a new consumable or joint preparation changed the answer, so each carries a stated range. Wu and Kim recommend inherent-strain models as the fast inner model for AI-driven optimisation, since thousands of full runs cost too much. See AI surrogate models you can validate.

From one distorted assembly to a distortion-control workflow you keep

The return comes when the next assembly, sequence or fixture revision runs through the same chain without a new model or new full-size trials.

• A seam library of calibrated heat-source and strain inputs per joint type, process, consumable and material, each with the measurement it was checked against and its pass-lumping rule.
• A weld-plan template in your existing solver that reads your weld list and takes sequence, direction and clamp steps as input, so a new sequence is a line in a file.
• A sequence study runner that executes the designed orders on your hardware, rejects failed runs and collects distortion at the interface points.
• A surrogate with a stated range and a gate: anything outside its joint types, grades and process window goes to the full model.
• A validation record against your first article, with tolerances written before the run and every model update logged.
• Handover and training, so your team adds seams itself. See our training and software development services.

Questions engineers ask about welding distortion simulation

If a weld sequence, fixture or pre-set is waiting on a distortion answer, send us the weld list (seams, process, passes, consumable, material), the drawing with its interface tolerances, your current sequence and fixture arrangement, and measured deviations of one finished assembly if you have them. We reply with how we would split the joint model from the assembly model, which seams to calibrate first, and what the validation record should hold. Start with a conversation.