RSFM: Reduced-Complexity Fluid Flows Database

• RSFM is a unified, open-access database that standardizes high-fidelity flow datasets and protocols for fair benchmarking of reduced-complexity fluid models.
• RSFM supports key capabilities, including compression, forecasting, and sensing, by providing datasets spanning laminar, turbulent, and experimental flows.
• RSFM ensures reproducibility and objective evaluation through baseline implementations, standardized metrics, and blind-testing protocols.

The Reduced-Complexity Modeling of Fluid Flows Database (RSFM) is a unified, open-access platform meticulously designed to facilitate the development, benchmarking, and fair comparison of data-driven and physics-based reduced-complexity models for a diverse array of fluid flow phenomena. RSFM stands at the intersection of contemporary computational science, machine learning, and fluid mechanics, addressing the urgent need for scalable, interpretable, and efficient surrogates capable of compression, forecasting, and sensing in canonical and application-specific flows. The database is structured to ensure reproducibility and rigorous assessment across methods, encapsulating reference datasets, standard metrics, baseline implementations, and blind-testing protocols (Schmidt et al., 7 Jan 2026, Towne et al., 2022).

1. Scope and Objectives

RSFM is architected to provide standardized, high-fidelity flow datasets for direct model benchmarking, supporting three primary capabilities:

• Compression: Creation of compact representations for very large flow datasets, facilitating storage, transmission, and real-time applications.
• Forecasting: Prediction of future flow states from a finite historical record, encompassing regimes from periodic laminar dynamics to high-Reynolds-number turbulence.
• Sensing: Inference of unmeasured flow states using limited or sparsely sampled measurements, crucial for experiment and control.

The database's principal aim is to enable equitable, objective, and reproducible comparisons between traditional modal techniques (e.g., POD, DMD), hybrid physics-ML models, and state-of-the-art deep learning surrogates (Schmidt et al., 7 Jan 2026, Towne et al., 2022).

2. Flow Datasets and Metadata Structure

RSFM organizes its contents around six flagship sub-challenges, each corresponding to distinct canonical or application-focused flow regimes. The datasets span laminar and turbulent, stationary and transient, two- and three-dimensional, simulated and experimental flows:

• 2D/3D Turbulent Boundary Layer DNS (BL1, BL2): Zero-pressure-gradient boundary layers, high spatial and temporal resolution, suitable for compression and near-wall sensing (Towne et al., 2022).
• Turbulent Jet LES: Axisymmetric Mach 0.9 jet, focusing on volumetric reconstruction and causal estimation from sparse probes.
• Turbulent Cavity Flow (TR-PIV): High-frequency experimental velocity measurements in an open cavity, relevant for short-horizon forecasting.
• Pitching Airfoil DNS: Parametric variation in pitching amplitude and frequency, emphasizing data-driven surrogate forecasting for laminar and transitional regimes.
• Experimental Gust–Airfoil Encounters: Combined PIV and force datasets for benchmarking control-oriented reduced models.
• Separated Airfoil Wake LES: High Reynolds number, three-dimensional complexity, challenging for both data-driven and physics-based ROMs.

All data is stored in HDF5 format with a uniform directory structure, explicit group/dataset naming, and detailed metadata (geometry, units, nondimensional numbers, spatial/temporal resolution, data provenance) (Towne et al., 2022).

3. Standardized Benchmarking Protocols

RSFM supplies scalar metrics, baseline implementations, and open-source evaluation tools to enforce cross-method comparability:

• Compression Metrics: Normalized mean-squared error (NMSE), compression ratio (CR), and relative wall-clock time, enabling assessment of fidelity vs. reduction (Schmidt et al., 7 Jan 2026).
• Forecasting Metrics: Time-dependent NMSE, quantifying error growth over prediction horizon for dynamic accuracy.
• Sensing Metrics: Instantaneous NMSE, evaluating the accuracy of reconstructed target fields from sparse measurements.

Baseline methodologies include space-only POD, convolutional autoencoders, Exact DMD, LSTM networks, linear stochastic estimation (LSE), Wiener filtering, MLPs, and CNNs, each available via example drivers and reference scripts (Schmidt et al., 7 Jan 2026).

Submissions are evaluated through blind-testing: only input or sensor measurements are released for test cases, with ground-truth withheld. Final metrics are computed strictly by the organizing committee via standardized scripts, mitigating potential overfitting and ensuring fairness (Schmidt et al., 7 Jan 2026).

4. Representative Reduced-Complexity Modeling Workflows

RSFM example workflows showcase representative approaches for each dataset:

• Spectral POD and Resolvent Analysis: Extraction of energetically dominant flow structures and frequency-specific modes for jet and boundary-layer data; resolvent mode comparison provides physical interpretability (Towne et al., 2022).
• Data-driven ROMs and Hybrid Surrogates: MLPs and CNNs for rapid field reconstruction; LSTM-augmented POD/DMD for nonlinear forecasting in unsteady or turbulent flows.
• Physics-constrained and filtered ROMs: Methodologies employing explicit spatial filtering and data-driven closure modeling, such as DDF-ROM and CDDF-ROM, are cataloged for general nonlinear PDE reductions (Xie et al., 2017, Mohebujjaman et al., 2018).
• Nonlinear Parametric Modeling: SINDy-based sparse regression and interpretable, symmetry-constrained low-dimensional ODEs accommodate complex phenomena and parameter sweeps (Deng et al., 2021, Oishi et al., 2023).
• Probabilistic Surrogates and Uncertainty Quantification: Gaussian-mixture neural networks yield both prediction and quantitative confidence estimates, providing uncertainty-aware reconstruction for high-dimensional flows (Fukami et al., 2020).

5. Example Applications and Impact

RSFM enables, for the first time, systematic benchmarking and acceleration of model development across a spectrum of practical domains:

• Aerospace and Aerodynamics: Acceleration and compression of DNS and LES databases supporting control and optimization in airfoil, jet, and boundary-layer flows.
• Hemodynamics and Physiology: Efficient surrogate modeling for arterial networks and modified Womersley equations, relevant for real-time clinical simulations (San et al., 2012).
• Environmental and Geophysical Flows: Surrogates for atmospheric boundary layers and sea-surface temperature prediction, with interpretability and quantifiable uncertainty (Díaz et al., 2023, Fukami et al., 2020).
• Experimental Data Reconstruction: Non-intrusive estimation and sensing with sparse multi-modal sensor inputs, enhancing experimental throughput and analysis (Schmidt et al., 7 Jan 2026).

By collecting methods, codebases, data, and evaluation protocols in a unified schema, RSFM fosters reproducible research and transparent model selection.

6. Limitations and Recommendations

RSFM's current scope prioritizes canonical flows under moderate Reynolds numbers (up to Reτ≈1000, Mach≤0.9), and high-quality DNS/LES/PIV datasets. The database does not cover supersonic jets, combusting flows, wall-heat transfer, or strongly three-dimensional compressible regimes (Schmidt et al., 7 Jan 2026). Best practices include explicit use of provided metadata descriptors, volume weighting for nonuniform grids, reproducible initial condition specification, and ensemble averaging for experiments (Towne et al., 2022).

Computational constraints may arise, especially for baseline deep learning models in 3D. Model-storage overheads are excluded from CR but should be considered for deployment in memory-limited environments.

7. Access, Usage, and Future Directions

RSFM data and tools are available via:

• Web portal: https://fluids-challenge.engin.umich.edu/
• Code repository: https://github.com/fluid-challenge/rsfm
• Data archives: Deep Blue Data (University of Michigan), Globus endpoint “DeepBlueData#kk91fk98z”

Licensing is CC-BY-4.0 for data and MIT for code; citation guidelines are provided. The community challenge model encourages negative results, thorough limitation analysis, and promotes open dissemination via AIAA Journal Virtual Collection and conferences (Schmidt et al., 7 Jan 2026).

Future expansion is anticipated to include a broader range of flow regimes, multimodal experimental and simulation data, and advanced uncertainty-aware metrics, reinforcing RSFM's role as a central resource for scalable, interpretable, and robust reduced-complexity modeling in fluid mechanics.