Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
FlowMixer: A Depth-Agnostic Neural Architecture for Interpretable Spatiotemporal Forecasting
Authors: Fares Mehouachi, Saif Eddin Jabari
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments across diverse domains demonstrate Flow Mixer s long-horizon forecasting capabilities while effectively modeling physical phenomena such as chaotic attractors and turbulent flows. Our results achieve performance matching state-of-the-art methods while offering superior interpretability through directly extractable eigenmodes. |
| Researcher Affiliation | Academia | Fares B. Mehouachi New York University in Abu Dhabi Abu Dhabi, UAE EMAIL Saif Eddin Jabari New York University Abu Dhabi, UAE & Brooklyn, USA EMAIL |
| Pseudocode | No | The paper describes the architecture and its components mathematically and textually in Section 3 and its subsections, but does not include a clearly labeled pseudocode or algorithm block. |
| Open Source Code | Yes | The code for Flow Mixer is available at https://github.com/Fares BMehouachi/Flow Mixer. |
| Open Datasets | Yes | We evaluate Flow Mixer on benchmark datasets widely used in the forecasting literature, including ETT (electricity transformer temperature), Weather, Electricity, and Traffic datasets [42]. To illustrate these patterns, we compute KK eigenmodes from the traffic dataset [46] enabling effective visualization of spatial components. |
| Dataset Splits | Yes | Data preparation followed standard protocols, with chronological partitioning using train/validation/test ratios of 0.6/0.2/0.2 for ETT and 0.7/0.2/0.1 for others. |
| Hardware Specification | Yes | Model training was implemented in Tensor Flow and executed on a single NVIDIA A100 GPU. |
| Software Dependencies | Yes | Model training was implemented in Tensor Flow and executed on a single NVIDIA A100 GPU. Implementation: Py Torch 2.0, NVIDIA A100 GPU, consistent random seeds for reproducibility |
| Experiment Setup | Yes | The training protocol incorporated early stopping with patience of 10 epochs and learning rate reduction (factor 0.1, patience 5) based on validation loss. We limited training to 100 epochs with batch size 32 and systematically evaluated dropout rates between 0.0 and 0.7. The model s hyperparameters were optimized for each dataset and prediction horizon combination, with comprehensive configurations detailed in Table 3. |