Springback Compensation

Academic Research / Computational Engineering

Uncertainty-aware iterative tool correction for deep-drawn sheet-metal components.

Noisy FE simulations per iteration
30
Iterations with α = 0.8
≈ 3
Iterations with α = 0.5
≈ 4
Iterative springback compensation workflowSemester thesis · original figure
Original figure from the public research repository.
Interactive study overview using the documented case-study configuration

Overview

An iterative FE and geometry-processing workflow that converts springback deviations into a corrected tool surface for the next forming simulation.

Semester thesis using FE ensembles, geometry reconstruction, and normal-direction surface compensation.

After a deep-drawn sheet is released from its tooling, elastic recovery changes the final geometry. For aluminium parts, this springback can create deviations large enough to prevent assembly.

The thesis investigated whether an ensemble of noisy FE results could drive a stable correction loop: reconstruct the deformed part, measure deviations along target-surface normals, and modify the tool in the opposite direction.

The problem

Directly compensating one simulated surface can amplify local noise and produce unstable tooling. The workflow therefore needed consistent geometry correspondence, robust ensemble aggregation, and safeguards against over-compensation near convergence.

Why this was difficult

01

Geometry correspondence

Every simulated STL had to be rigidly aligned, trimmed, sampled, and reconstructed onto a comparable surface representation.

02

Uncertain simulation response

Thirty noisy FE results per iteration produced a distribution of possible compensated positions rather than one deterministic answer.

03

Convergence versus stability

Aggressive correction reduces large initial errors quickly but can bend the tool excessively when residual errors become small.

My contribution

Developed the research pipeline for mesh alignment, structured surface reconstruction, ensemble aggregation, normal-deviation calculation, localized compensation weighting, and STEP export for iterative AutoForm evaluation.

System workflow

  1. 01Align the FE ensemble

    Register 30 springback meshes to the target using the part-bottom reference region.

  2. 02Reconstruct comparable surfaces

    Trim the flange and approximate each result with a B-spline surface using shared parameterization.

  3. 03Reduce uncertainty

    Aggregate corresponding control points using mean, median, or maximum target deviation.

  4. 04Calculate compensation

    Project deviations along target normals and apply global alpha and local beta weighting.

  5. 05Generate and evaluate the next tool

    Displace control points inversely, export STEP geometry, and run the next AutoForm iteration.

Technical decisions

Decision 01

Measure deviations along target-surface normals

geometric metric
Problem
Raw nearest-point distances do not preserve a stable correction direction on curved sheet geometry.
Decision
Evaluate signed deviation along target normals and use that direction for inverse tool displacement.
Result
Compensation remains connected to the nominal surface geometry and local forming direction.
Decision 02

Aggregate an ensemble before compensating

uncertainty handling
Problem
One noisy FE result can push control points toward a non-representative tool correction.
Decision
Compare arithmetic mean, coordinate-wise median, and maximum-deviation reductions over 30 simulations.
Result
The study exposes the trade-off between robust convergence and aggressive early correction.
Decision 03

Treat the maximum strategy as phase-dependent

stability control
Problem
The farthest simulated point reduces large deviations quickly but can over-compensate close to convergence.
Decision
Analyze aggressive maximum reduction separately from stable mean/median behavior and propose switching strategy by iteration phase.
Result
Failure behavior becomes part of the engineering conclusion rather than being hidden by one aggregate score.

Results

These results come from separate experiments and model configurations. They should be interpreted individually, not as the performance of one combined final model.

Compensation A≈ 3 iterationsTolerance convergence
Variant
α = 0.8
Evaluation
Included aluminium deep-drawing case
Compensation B≈ 4 iterationsTolerance convergence
Variant
α = 0.5
Evaluation
Included aluminium deep-drawing case
Robustness studymean / median / maxEnsemble-reduction strategies
Evaluation
30 noisy simulations per iteration

These observations describe the included thesis case study and are not a general guarantee for other materials, meshes, geometries, or AutoForm settings.

Lessons & takeaways

  • The fastest correction strategy is not necessarily the most stable strategy near convergence.
  • Reliable geometry compensation depends on correspondence and reference alignment before deviation statistics are meaningful.
  • A phase-dependent strategy—aggressive early, robust late—is a promising extension of the observed behavior.

Limitations

  • AutoForm is required to reproduce the complete FE loop; the repository provides code and reference data around that external solver workflow.
  • The reported convergence behavior is specific to the documented aluminium demonstrator and study configuration.
Next project17CellVisionAcademic / Computer Vision

Project image

Interactive demo