Verifier-Guided Model Discovery Enhances Symbolic Transformers for Physical Dynamics
A new verifier-guided (VG) workflow has been created by researchers to enhance the transferability of pretrained symbolic transformers when applied to high-dimensional physical data, tackling a significant challenge in scientific discovery. This method, which utilizes the ODEFormer backbone, incorporates dynamical and physical admissibility criteria to choose from a pool of candidate equations with multiple trajectories. In experiments involving canonical Van der Pol oscillators, the VG workflow surpassed the original ODEFormer across various held-out initial conditions. Additionally, the technique was tested on vortex shedding in fluid dynamics, showcasing its applicability to a wider range of physical systems. The findings are presented in a paper on arXiv (2608.02662), emphasizing the potential of integrating machine learning with physical constraints for improved generalizability and interpretability in scientific modeling.
Key facts
- Verifier-guided (VG) workflow enhances transfer of pretrained symbolic transformers to high-dimensional physical data.
- The workflow uses ODEFormer as a symbolic backbone.
- Selection criteria include dynamical and physical-admissibility.
- VG outperforms original ODEFormer on Van der Pol oscillators across held-out initial conditions.
- Application to vortex shedding is addressed.
- The paper is available on arXiv with identifier 2608.02662.
- The approach aims to improve forecasting of nonlinear physical systems.
- The method promises interpretable alternatives to opaque machine-learning surrogates.
Entities
Institutions
- arXiv