Divergence-aware adaptive prediction framework for accelerating CFD simulations of unsteady flows

Abstract: Reliable long-horizon prediction remains a challenge for data-driven CFD surrogates, because offline-trained models accumulate autoregressive errors and lose accuracy when operating conditions change. This work develops a divergence-aware adaptive CFD-surrogate framework that couples a CFD solver with a proper orthogonal decomposition-deep learning (POD-DL) surrogate in a closed-loop workflow. CFD snapshots are compressed by POD, and a neural-network predictor advances the reduced state in time. The surrogate performs autoregressive forecasting, while its reliability is monitored online. When a prescribed update interval is reached or prediction degradation is detected, the CFD solver is automatically recalled to generate new snapshots and update the surrogate. The framework is assessed for three-dimensional flow past a circular cylinder at Re = 160-400. Baseline non-adaptive predictions exhibit progressive error growth over long forecast horizons, confirming the need for online correction. With prescribed update intervals, the adaptive framework preserves the dominant wake dynamics and reduces error growth after retraining compared with the non-adaptive model. For a representative 200-snapshot interval, the framework achieves a speed-up ratio of approximately 92 relative to CFD. An event-triggered mode is introduced using ensemble uncertainty and dynamically estimated thresholds. This mode terminates unreliable forecasts without requiring ground-truth CFD data during prediction, and the detected triggers are consistent with the onset of deterioration in the lift-coefficient evolution. Under varying inlet conditions, the framework detects regime changes, recalls CFD, and recovers reliable predictions. These results demonstrate that divergence-aware CFD-surrogate coupling provides a robust and efficient route for adaptive long-horizon flow prediction under evolving operating conditions.