Verifier-Guided Model Discovery for Physical Dynamical Systems with Pretrained Symbolic Transformers

Short summary

This paper presents a verifier-guided workflow around ODEFormer that uses dynamical and physical-admissibility criteria to select candidate equations from a symbolic transformer, enabling transfer to high-dimensional physical data. On Van der Pol oscillators and vortex shedding problems, the approach discovers interpretable reduced-order equations without system-specific libraries. The authors find that reconstruction fidelity alone doesn't determine symbolic discoverability, emphasizing backbone pretraining distribution compatibility.

• •Verifier-guided workflow wraps ODEFormer with physical-admissibility checks for symbolic equation discovery
• •Successfully discovers reduced-order equations for vortex shedding without prescribed Navier-Stokes structure
• •Reconstruction fidelity alone doesn't guarantee symbolic discoverability — backbone compatibility matters

Generated with AI, which can make mistakes.