Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

U Net LSTM with incremental time-stepping for robust long-horizon unsteady flow prediction

Blaise Madiega, Mathieu Olivier

TL;DR

This work tackles the high computational cost of long-horizon unsteady CFD by introducing an Incremental Time-Stepping U‑Net–LSTM that learns per-step field updates (Δφ) rather than absolute states. By predicting increments and updating states as φ_t ≈ φ_{t−1} + Δφ_t, the approach reduces autoregressive drift and aligns with classical time marching, enabling smoother long-term predictions and potential integration as a predictor or corrector in CFD solvers. Across three test cases (2D cylinder wake, 2D Kelvin–Helmholtz shear layer, and 3D turbulent channel flow), the incremental model achieves 54.5%–84.2% reductions in cumulative error for long-horizon rollouts, while preserving key engineering metrics; it also dramatically reduces memory needs (≈35×) at the expense of some training time and minor accuracy loss. These findings support a practical, scalable path to hybrid CFD–ML pipelines that accelerate simulations without sacrificing quantitative fidelity, with future work focused on fully embedding the predictor into PISO/PIMPLE loops and exploring adaptive time stepping and broader generalization.

Abstract

Transient computational fluid dynamics (CFD) remains expensive when long horizons and multi-scale turbulence are involved. Data-driven surrogates promise relief, yet many degrade over multiple steps or drift from physical behavior. This work advances a hybrid path: an incremental time-stepping U Net LSTM model that forecasts unsteady dynamics by predicting field updates rather than absolute states. A U-Net encoder decoder extracts multi-scale spatial structures, LSTM layers carry temporal dependencies, and the network is trained on per-step increments of the physical fields, aligning learning with classical time marching and reducing compounding errors. The model is designed to slot into solvers based on projection methods (such as SIMPLE, PISO, etc), either as an initializer that delivers a sharper first guess for pressure-velocity coupling or as a corrective module that refines provisional fields. Across representative test cases, the approach improves long-term stability (54.53 to 84.21 % reduction of cumulative errors) and preserves engineering metrics, integral and averaged quantities, more reliably than standard learning baselines. These properties make it a plausible component of hybrid CFD-ML pipelines designed to accelerate unsteady simulations without compromising quantitative fidelity.

U Net LSTM with incremental time-stepping for robust long-horizon unsteady flow prediction

TL;DR

This work tackles the high computational cost of long-horizon unsteady CFD by introducing an Incremental Time-Stepping U‑Net–LSTM that learns per-step field updates (Δφ) rather than absolute states. By predicting increments and updating states as φ_t ≈ φ_{t−1} + Δφ_t, the approach reduces autoregressive drift and aligns with classical time marching, enabling smoother long-term predictions and potential integration as a predictor or corrector in CFD solvers. Across three test cases (2D cylinder wake, 2D Kelvin–Helmholtz shear layer, and 3D turbulent channel flow), the incremental model achieves 54.5%–84.2% reductions in cumulative error for long-horizon rollouts, while preserving key engineering metrics; it also dramatically reduces memory needs (≈35×) at the expense of some training time and minor accuracy loss. These findings support a practical, scalable path to hybrid CFD–ML pipelines that accelerate simulations without sacrificing quantitative fidelity, with future work focused on fully embedding the predictor into PISO/PIMPLE loops and exploring adaptive time stepping and broader generalization.

Abstract

Transient computational fluid dynamics (CFD) remains expensive when long horizons and multi-scale turbulence are involved. Data-driven surrogates promise relief, yet many degrade over multiple steps or drift from physical behavior. This work advances a hybrid path: an incremental time-stepping U Net LSTM model that forecasts unsteady dynamics by predicting field updates rather than absolute states. A U-Net encoder decoder extracts multi-scale spatial structures, LSTM layers carry temporal dependencies, and the network is trained on per-step increments of the physical fields, aligning learning with classical time marching and reducing compounding errors. The model is designed to slot into solvers based on projection methods (such as SIMPLE, PISO, etc), either as an initializer that delivers a sharper first guess for pressure-velocity coupling or as a corrective module that refines provisional fields. Across representative test cases, the approach improves long-term stability (54.53 to 84.21 % reduction of cumulative errors) and preserves engineering metrics, integral and averaged quantities, more reliably than standard learning baselines. These properties make it a plausible component of hybrid CFD-ML pipelines designed to accelerate unsteady simulations without compromising quantitative fidelity.

Paper Structure

This paper contains 22 sections, 6 equations, 15 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Dataset Generation
  4. Methodology
  5. Dataset Generation
  6. Training Processes
  7. U-Net-LSTM Architecture
  8. Increments Prediction Approach
  9. Incremental Learning
  10. Experimentation
  11. Implementation notes
  12. Data windowing and targets
  13. Model and training protocol
  14. Hyperparameters' optimization with Optuna
  15. Data splits and metrics
  16. ...and 7 more sections

Figures (15)

  • Figure 1: U– Net--LSTM architecture: U– Net for multiscale spatial feature extraction and LSTM for sequence modeling in unsteady flow analysis.
  • Figure 2: Schematic illustration of the two prediction approaches: (left) the classic approach, where the model U-Net-LSTM (illustrated at Figure \ref{['fig:unet_lstm_architecture']}) predicts absolute values of physical fields; (right) the incremental prediction approach, where the model predicts velocity field increments at each timestep.
  • Figure 3:
  • Figure 4:
  • Figure 5:
  • ...and 10 more figures