GeoMatries: Machine Learning-Empowered Simulation for 4D Printing and Morphing Materials Design

Materials in nature are mostly active, responsive, and transformative. Yet in conventional design practices,
these features are often neglected and removed from the final products. By contrast, morphing matter design leverages the tunable, temporal properties of materials to program functions into artifacts, and requires accurate physical performance predictions to inform design decisions. 4D printing, in particular, is an additive manufacturing technique that manipulates residual stress to create stimuli-responsive, shape-changing artifacts. However, due to the lack of a fast and physically accurate simulation method to
inform design decisions, the current design workflow of 4D printing requires intensive physical prototyping to iterate designs, making it a slow, inefficient, and indirect process.
This thesis takes 4D printing as an example of morphing matter design and responds to the workflow challenges mentioned above with SimuLearn, a data-driven simulation technique that combines numerical methods and machine learning, to mitigate the need of physical prototyping in design iterations. Compared to finite element analysis, our results show that this technique can simulate physical
transformations with speed (1500 times faster) while having an identical accuracy (1.6% maximum relative error). Workflows adopting SimuLearn will be able to afford interactive and physically-informed digital iterations and extend the design space of 4D printing. Additionally, a prototype computer-aided morphing matter design tool is also implemented to expose the development guidelines of tools that adopt SimuLearn and is deployed in several design tasks to demonstrate its applicability and potential.
Lastly, this thesis will also discuss the limitations, generalizability, and potential improvements of SimuLearn to guide future works and real-world deployments.