Student projects

The harder and largely unsolved step comes earlier: deciding how an object will actually be used, and therefore loaded, then reworking its internal structure rather than its outline so that it stays both recognizable and sound. The project builds a pipeline that reasons about an object and its grasp, contact, and support regions to propose physically meaningful load cases on its own and then runs structural optimization that is free to change topology. Selected results are fabricated and compared against baselines.

Deliver a system that turns a generated or scanned 3D object, together with a short description of what it is for, into a set of inferred load cases and an optimized, fabricable structure, and quantify when inferred loads and genuine topology change outperform manual, surface-only baselines. **Deliverables of this project include:** - A module that infers plausible load cases from an object and its function. - Integration with a three-dimensional structural optimizer that is free to change topology. - A benchmark comparison against surface-only and hand-specified baselines. - Fabricated demonstrators of selected results. **Desirable student qualities:** - Strong programming and machine learning skills in Python and PyTorch. - Familiarity with modern 3D representations such as implicit fields or Gaussian Splatting is a plus. - Some background in mechanics or finite element analysis, or a clear willingness to learn it. – Interest in generative and language models for physical design and good knowledge of English.

Dr. Tino Stankovic tinos@ethz.ch Dr. Amira Abdel-Rahman abdela@ethz.ch / amiraa@cornell.edu

Most current methods sidestep this by training an offline surrogate of the micro response before optimizing, which is costly and tied to the sampled design space. The project instead resolves the nonlinear micro response directly inside a fully differentiable optimization loop, so that geometry at both scales is driven by large deformation targets, with the strongest designs validated by highresolution elastomer printing.

Build a differentiable two-scale optimization pipeline that designs soft architected materials for prescribed nonlinear targets, such as a tailored force-displacement curve or an energy absorption profile, without a precomputed surrogate, and demonstrate it on benchmark cases with printed and tested samples. **Deliverables of this project include:** - A differentiable large-deformation simulation and homogenization module. - A two-scale optimization loop running on the GPU. – Benchmark designs for at least two distinct nonlinear targets. - Fabricated samples with a comparison between predicted and measured response. **Desirable student qualities:** - Background in solid and continuum mechanics and finite element analysis. - Strong programming, ideally with an automatic differentiation framework such as JAX, PyTorch or Taichi. - Interest in optimization and in linking simulation to fabrication. - Good knowledge of English.

Dr. Tino Stankovic tinos@ethz.ch Dr. Amira Abdel-Rahman abdela@ethz.ch / amiraa@cornell.edu