MechE-CSE PhD Thesis Defense: Sanghee Lee
Abstract:
Hyperelastic constitutive modeling provides a framework for analyzing large elastic deformations of rubber-like solids, foams, and other soft materials. These materials exhibit diverse nonlinear responses that are often difficult to describe with prescribed analytical strain-energy functions, which motivates representations learned from experimental data. Machine learning offers new ways to construct such representations and is also being used to develop procedures for solving mechanics problems. These developments raise two related questions about what a learned model can represent and whether an implemented model or procedure is reliable. This thesis addresses these questions by developing flexible and interpretable constitutive models for incompressible and compressible isotropic hyperelasticity, together with practical basic tests for assessing the mechanical behavior of learned models and solution procedures.
We first develop patch tests for machine-learning solution procedures. The tests are carried out with the procedure itself, without introducing any separate analysis method. The tests require the recovery of rigid-body motions and constant strain and stress states, for which the correct responses are known. In an illustrative physics-informed neural network, a deliberately introduced implementation error causes the procedure to fail the tests, even though the training losses decrease and some benchmark results appear plausible. This example shows that loss reduction and a plausible structural response can be insufficient evidence of mechanical correctness.
Next, we develop a trainable cubic B-spline representation for incompressible isotropic hyperelasticity within the Sussman–Bathe formulation. The spline coefficients and interior knot locations are learned jointly, allowing the model to follow local variations in the material response without interpolating every fluctuation in the data. After training, the spline provides an explicit material law that can be evaluated directly in finite element analysis. The trained model reproduces the analytical stress states in constitutive patch tests.
The model is then extended to compressible materials. We introduce a learnable coupling parameter that controls how much volumetric strain enters the principal logarithmic strain coordinates. The principal and volumetric energy derivatives are built on the learnable B-spline architecture developed for the incompressible model. We conduct parameter recovery tests on synthetic data and apply the model to data from three open-cell elastomeric foams of different densities that exhibit pronounced tension–compression asymmetry. The explicit structure of our model allows prior knowledge of the similarity among the foams to be incorporated through shared knots.
A close stress fit at measured states does not guarantee the correct deformation when the strains are determined through equilibrium. We therefore assess the learned laws not only by using finite element analysis to predict shear responses excluded from training, but also by reconstructing lateral strains of the foams under traction-free uniaxial loading. The predicted shear responses are comparable in accuracy to those of other models, and the reconstructed strains closely follow the measured values, providing a constitutive consistency check along the training paths. Together, the proposed models and tests show how constitutive modeling can gain flexibility while remaining interpretable and subject to practical basic checks of its mechanical behavior.
Thesis Committee Members:
- Professor Klaus-Jürgen Bathe, Professor Emeritus of Mechanical Engineering, MIT (Thesis Advisor)
- Professor David M. Parks, Professor of Mechanical Engineering, MIT
- Professor Markus J. Buehler, Jerry McAfee (1940) Professor in Engineering, MIT