Stable Fine-Time-Step Long-Horizon Turbulence Prediction with a Multi-Stepsize Mixture-of-Experts Neural Operator
Stable Fine-Time-Step Long-Horizon Turbulence Prediction with a Multi-Stepsize Mixture-of-Experts Neural Operator
Authors: Guanyu Pan, Huiyu Yang, Yunpeng Wang, Zikun Xu, Jianchun Wang, Nianyu Yi Date: 2026-04-14 Paper ID: openalex:2604.12794
Summary
This paper addresses the stability issues in long-horizon autoregressive neural operator predictions for turbulent flows, which are often caused by error accumulation at fine temporal resolutions. The authors propose a multi-stepsize mixture-of-experts (Ms-MoE) framework built on an implicit factorized Transformer (IFactFormer) backbone that adapts to different requested relative strides. By utilizing a time-step router to select scale-specific experts, the model maintains high-fidelity predictions during long-term rollouts. Empirical results on homogeneous isotropic turbulence and turbulent channel flow demonstrate superior stability and statistical accuracy over existing methods.
Key Contributions
- Introduces the Multi-stepsize Mixture-of-Experts (Ms-MoE) neural operator that enables stable long-horizon turbulence forecasting at fine temporal resolutions.
- Proposes the Implicit Factorized Transformer (IFactFormer) as a backbone architecture for modeling time-advancement operators.
- Demonstrates improved stability and agreement with long-time-averaged statistics on forced homogeneous isotropic turbulence (HIT) and turbulent channel flow datasets compared to standard autoregressive baselines.
Open Questions & Future Work
Key Concepts
- ms-moe-neural-operator: A mixture-of-experts architecture that conditions on temporal stride to maintain stability and accuracy across varying time-steps in forecasting models.
Archivist Review
I have approved the Ms-MoE concept as a reusable architecture for temporal stride conditioning. I also approved a more generalized version of the open question regarding stability and compositional consistency in neural operators, as this represents a fundamental challenge in scientific ML time-series forecasting. The proposed open question from the analysis was rejected in favor of the newly defined one to better align with the vault’s standards for tracking long-term research bottlenecks.
Approved Concepts
- Multi-stepsize Mixture-of-Experts (Ms-MoE): Provides a novel mechanism for handling multi-scale temporal dynamics in neural operators, which is a common bottleneck in long-horizon forecasting.
Approved Open Questions
- Long-horizon Neural Operator Stability: Crucial for the long-term reliability of surrogates in scientific computing where statistical accuracy is as important as point-wise error.
Rejected Candidates
- [open_question] Stable Long-Horizon Turbulence Prediction (
stable-long-horizon-turbulence-prediction) - duplicate_existing: Duplicate of or subsumed by the more general and better-defined ‘long-horizon-neural-operator-stability’.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.12794
- paper_id
- 2604.12794
- paper_source
- openalex
- domain
- time-series
- tags
- time-seriesforecasting
- architectures
-
- datasets
-
- concept_slugs
- ms-moe-neural-operator
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-17T05:47:03Z