Research paper analyzes optimal input sizing for neural field forecasting

A new research paper explores the optimal input configuration for neural surrogates used to forecast physical fields. The study, which analyzes the trade-off between spatial context (window size) and temporal memory (history length), found that for fields with local dynamics, the optimal forecast relies more on spatial correlation than temporal history. The research suggests that for surrogate design, the input window should be sized according to the field's correlation lengths, and the input budget should be split between window and history based on the model's intended lead time, as the optimal split shifts towards wider windows with increasing lead times. AI

IMPACT Provides insights into optimizing neural network architectures for forecasting physical phenomena, potentially improving efficiency in scientific modeling.

RANK_REASON The cluster contains a single academic paper detailing novel research findings. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Bayes-optimal error
• Burgers flow
• Hugging Face
• Kuramoto-Sivashinsky (KS) equation
• Lyapunov exponent
• Neural surrogates

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →