Deep Wave Network architecture improves accuracy-cost trade-off for physical dynamics modeling

Researchers have introduced the Deep Wave Network (DW-Net), an architectural innovation for U-Net-type models used in physical dynamics modeling. DW-Net enhances effective depth by stacking multiple encoder-decoder "waves" in series, incorporating skip connections both within and across these waves for progressive refinement. This approach consistently improves the accuracy-cost trade-off compared to traditional single-wave U-Nets, achieving higher accuracy at matched computational cost or similar accuracy with reduced training time, as demonstrated across various 2D and 3D flow benchmarks. AI

影响 Introduces a new architecture that improves the accuracy-cost trade-off for physical dynamics modeling, potentially leading to more efficient simulations.

排序理由 This is a research paper detailing a novel neural network architecture for physical dynamics modeling. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.LG 阅读 →

• arXiv
• Deep Wave Network
• U-Net

• 论文
• 其他

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →