DEFT: Data-Efficient Frequency-domain Top-k Sampling via Inverse Discrete Fourier Transform for Spatiotemporal Dynamical Systems Modeling
DEFT generates training data by perturbing dominant Fourier modes, then reconstructing physically consistent samples with an inverse DFT.
The paper targets PDE-driven spatiotemporal systems where simulators are costly and data-heavy models can fail under changing dynamics. In PDEBench tests on diffusion-sorption and Burgers equations, the authors report a 40% cut in data needs with under 2% loss in predictive accuracy. In a battery degradation PDE system, DEFT reaches R² above 0.99 across test datasets, and its frequency-domain features transfer to other battery chemistries with 20% of the fine-tuning data. Source: ArXiv · AI/CL/LG's note.
The paper targets PDE-driven spatiotemporal systems where simulators are costly and data-heavy models can fail under changing dynamics. In PDEBench tests on diffusion-sorption and Burgers equations, the authors report a 40% cut in data needs with under 2% loss in predictive accuracy. In a battery degradation PDE system, DEFT reaches R² above 0.99 across test datasets, and its frequency-domain features transfer to other battery chemistries with 20% of the fine-tuning data. Source: ArXiv · AI/CL/LG's note.
score 4