Zhu et al. (2026) AnytimeHydroTwin: A deployment-valid inference-orchestration layer for AI-driven hydrological digital twins

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-09-24
• Authors: Yue Zhu, Qingyang Liu
• DOI: 10.1016/j.jhydrol.2026.136485

Research Groups

• College of Computing, Georgia Institute of Technology
• Khoury College of Computer Sciences, Northeastern University

Short Summary

This paper introduces AnytimeHydroTwin, an inference-orchestration layer for AI-driven hydrological digital twins. The layer combines a high-fidelity (HF) neural model with lag-1 persistence on a per-timestep basis and determines when to invoke the HF model or defer to a cheap persistence baseline.

Objective

• Investigate the accuracy-compute trade-off in hydrological digital twin operating loops
• Develop an inference-orchestration layer that can adaptively allocate compute resources between HF models and low-fidelity (LF) persistence predictions

Study Configuration

• Spatial Scale: Tidal-estuarine systems, including Sacramento Delta, Delaware basin, and Chesapeake-tributary Pamunkey
• Temporal Scale: 15-minute to hourly cadence, with a focus on short-horizon nowcasting

Methodology and Data

• Models used:
  • High-fidelity (HF) neural models (LSTM, TCN, Transformer)
  • Low-fidelity (LF) lag-1 persistence model
• Data sources:
  • USGS National Water Information System (NWIS) water-quality records
  • CAMELS basins dataset for cross-basin generalization

Main Results

• The inference-orchestration layer successfully adapts to the signal and allocates compute resources between HF models and LF persistence predictions, achieving a balance between accuracy and compute efficiency.
• The layer demonstrates improved performance over traditional HF-only approaches in certain regimes, while deferring to lag-1 persistence in others where it is sufficient.

Contributions

• Introduce an inference-orchestration primitive for hydrological digital twins that combines HF models with LF persistence predictions on a per-timestep basis.
• Develop a deployment-valid pre-HF causal gate that ensures the reported HF-invocation rate is a real compute saving, not a post-hoc blend ratio.

Funding

• This research was supported by [project/program name] (reference code: [code]) and [project/program name] (reference code: [code]).

Citation

@article{Zhu2026AnytimeHydroTwin,
  author = {Zhu, Yue and Liu, Qingyang},
  title = {AnytimeHydroTwin: A deployment-valid inference-orchestration layer for AI-driven hydrological digital twins},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136485},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136485}
}