On the feasibility of foundational models for neural simulation
Please login to view abstract download link
We investigate the potential of foundation models for simulating physical phenomena, focusing specifically on continuum mechanics (both solid and fluid). While "learned simulators" have shown promise in specific applications, it is still unclear how well they can handle significant changes in domain shape, boundary conditions, and/or constitutive laws while maintaining robustness (i.e., avoiding hallucinations) and accuracy. In this work, we conduct a comprehensive study to examine these capabilities, deliberately placing the models in challenging scenarios to assess their performance under severe alterations in their application domain.