Material Model
Academic Research / Scientific Machine Learning
Inverse and neural identification of anisotropic sheet-material parameters from a biaxial specimen.
Overview
A FEM–ML workflow for identifying YLD2000-2d material parameters from experimental full-field strain responses.
Master's thesis combining experimental mechanics, finite-element simulation, optimization, and machine learning.
Sheet materials exhibit direction-dependent yielding that cannot be captured reliably with a single isotropic stress value. The thesis investigated how an innovative biaxial specimen can expose several stress states within one experiment and provide richer evidence for parameter identification.
The research connected measured full-field strains with LS-DYNA simulations and compared two inverse routes: iterative parameter optimization and a learned mapping trained on randomized simulation cases.
The problem
Material calibration is an ill-posed inverse problem: multiple parameter combinations can produce similar responses, while each high-fidelity simulation is expensive and experimental measurements contain spatial variation.
Why this was difficult
Coupled physical parameters
Directional yield stresses and r-values influence the strain field together, making isolated one-variable calibration insufficient.
Experiment-to-simulation alignment
Comparable response quantities had to be extracted consistently from measured and simulated strain fields.
Expensive inverse search
Repeated FE evaluations motivate a learned inverse model, but the training database must still cover the physically relevant parameter space.
My contribution
Designed and implemented the thesis workflow from simulation-data generation and response extraction to LS-OPT calibration, neural model training, inference, and comparison of identified material parameters.
System workflow
- 01Measure the specimen
Capture spatial strain responses from the innovative biaxial specimen under controlled loading.
- 02Build the FE counterpart
Represent the specimen and YLD2000-2d material law in LS-DYNA.
- 03Generate response data
Run randomized parameter sets and extract comparable x/y strain quantities.
- 04Identify parameters
Use LS-OPT inverse analysis and a trained neural network as complementary inverse routes.
- 05Validate the response
Re-simulate identified parameters and compare predicted strain behavior with the experiment.
Technical decisions
Use a biaxial specimen with spatially varied stress states
experimental design- Problem
- Conventional tests provide only a limited subset of the evidence needed for anisotropic calibration.
- Decision
- Use one specimen geometry that creates regions approaching uniaxial and biaxial loading, then evaluate full-field strains.
- Result
- A single experiment contributes several informative response regions to the inverse problem.
Compare iterative and learned inverse mappings
hybrid identification- Problem
- Optimization is interpretable but simulation-heavy; direct neural inference is fast but depends on synthetic-data coverage.
- Decision
- Implement both routes against the same parameterization and response representation.
- Result
- The study can compare their trade-offs instead of treating either method as universally superior.
Validate through forward simulation
physics validation- Problem
- A plausible parameter vector is not sufficient evidence that the material response has been identified.
- Decision
- Return inferred parameters to the FE model and compare the resulting strain response with experimental measurements.
- Result
- Evaluation remains tied to observable mechanics rather than parameter values alone.
Research scope
- 8
- YLD2000-2d parameters identified
DefinitionDirectional stresses, r-values, biaxial response, and exponent M
- 2
- Identification strategies compared
DefinitionIterative inverse analysis and neural-network-assisted inference
Lessons & takeaways
- Scientific ML is strongest when the learned inverse model remains connected to a forward physical simulation.
- Response representation and experimental design can matter as much as neural-network architecture.
- Optimization and machine learning offer complementary accuracy, cost, and interpretability trade-offs.
Limitations
- The repository documents a thesis research workflow, not a production calibration service.
- Conclusions are tied to the investigated specimen, parameter ranges, material law, and available experimental data.