แนวคิดหลัก
DPOT introduces a novel auto-regressive denoising pre-training strategy and scalable model architecture based on Fourier attention for large-scale PDE pre-training, achieving state-of-the-art results on diverse downstream tasks.
สถิติ
우리는 0.5B 매개변수로 10개 이상의 PDE 데이터셋에서 DPOT 모델을 사전 훈련했습니다.
DPOT은 다양한 PDE 벤치마크 및 하류 PDE 작업에서 최첨단 결과를 달성했습니다.
คำพูด
"DPOT introduces a new auto-regressive denoising pre-training strategy by injecting noise into training data."
"Extensive experiments show that DPOT achieves SOTA on various benchmarks and downstream tasks."