New HERO training method boosts long-horizon accuracy for neural operators

Researchers have introduced HERO (History-Enriched Rollout Training), a novel method designed to improve the long-horizon accuracy of autoregressive neural operators. This technique addresses the issue of error accumulation in models that recursively use their own predictions, a common problem when simulating time-dependent partial differential equations. HERO enhances standard training by incorporating relative supervision derived from the model's optimization history, using a ranked set of candidate rollouts to establish a more informative comparison baseline. Experiments across nine PDE benchmarks demonstrated that HERO consistently improves long-horizon accuracy and robustness without increasing inference time. AI

IMPACT Improves long-horizon accuracy and robustness for autoregressive models in scientific simulations.

RANK_REASON The cluster contains a research paper detailing a new training method for neural operators. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• autoregressive model
• HERO
• History-Enriched Rollout Training
• Neural Operators
• partial differential equations
• spectroscopy

• paper
• model release

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