Uncertainty-driven phase fields for hybrid constitutive models

See the publication: Mixing data-driven and physics-based constitutive models using uncertainty-driven phase fields.

This work was performed as part of my PhD, in collaboration with my advisors, Iuri Rocha and Frans van der Meer, as well as in collaboration with WaiChing Sun, whom I got to visit at Columbia University for three months.

Summary

Surrogate constitutive models in computational mechanics require a lot of data to be accurate, and this data is expensive to compute. In this work, we use a surrogate model (that is insufficiently trained) when possible, but locally revert back to the high-fidelity ground-truth model when necessary. This way, accurate simulations are maintained, while the use of the surrogate does lead to real speedups in computation time. Switching between the surrogate and ground-truth model can cause solver instabilities. Therefore, we use a phase-field to smoothen this transition.