Observation-Guided Neural Surrogate Learning for Scientific Simulation Emulation: A Single-Gauge Flood-Inundation Proof of Concept
Observation-Guided Neural Surrogate Learning for Scientific Simulation Emulation: A Single-Gauge Flood-Inundation Proof of Concept
Authors: Marzieh Alireza Mirhoseini Date: 2026-04-28 Paper ID: openalex:2604.25890
Summary
This paper presents an observation-guided neural surrogate learning framework to emulate high-resolution hydrodynamic simulations for urban flood-inundation mapping. The approach uses a coarse ensemble Gaussian-process/local analogue (EnsCGP) model to generate initial estimates and uncertainty proxies, which are subsequently refined by a U-Net-ASPP neural corrector. By leveraging sparse point-wise gauge data as a training constraint alongside simulation-derived inputs, the model achieves high accuracy in replicating complex flood dynamics while maintaining consistency with sparse real-world sensor observations.
Key Contributions
- Introduces an observation-guided surrogate framework that uses single-gauge records to constrain neural emulator training for high-resolution flood mapping.
- Develops a hybrid architecture combining an ensemble Gaussian-process/local analogue (EnsCGP) coarse estimator with a U-Net-ASPP neural corrector.
- Demonstrates R^2 ~ 0.99 and MAE < 0.01m on LISFLOOD-FP simulation benchmarks for urban flood-inundation emulation.
Open Questions & Future Work
Key Concepts
- observation-guided-neural-surrogate-learning: A framework that uses sparse point-wise observations to guide the training of a neural surrogate model for spatial simulation emulation.
Archivist Review
I have approved the core methodological concept of observation-guided surrogate learning and the critical open question regarding the validation protocols for such hybrid frameworks. These items address the specific challenge of aligning high-fidelity simulations with sparse, real-world sensor data, a key bottleneck in scientific machine learning. Other potential candidates, such as the specific architecture (EnsCGP) or internal model components, were rejected as they are implementation-specific and represent subcomponents rather than distinct, reusable paradigms.
Approved Concepts
- Observation-Guided Neural Surrogate Learning: Provides a novel paradigm for bridging sparse real-world sensor data with dense simulation outputs in scientific surrogate modeling.
Approved Open Questions
- Validation of Observation-Guided Surrogates: Addressing these concerns is crucial to verify that the observation-guided correction improves actual model skill rather than merely overfitting to historical gauge records, and to establish the robustness and scalability of such hybrid modeling architectures for operational applications.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.25890
- paper_id
- 2604.25890
- paper_source
- openalex
- domain
- time-series
- tags
- time-series-augmented-generationuncertainty-weighted-mean-squared-error
- architectures
-
- datasets
-
- concept_slugs
- observation-guided-neural-surrogate-learning
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-01T06:05:12Z