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..
Cloud4D: Estimating Cloud Properties at a High Spatial and Temporal Resolution
Authors: Jacob Lin, Edward Gryspeerdt, Ronald Clark
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Across a two-month deployment comprising six skyward cameras, our system delivers an order-of-magnitude improvement in space-time resolution relative to state-of-the-art satellite measurements, while retaining single-digit relative error (< 10%) against collocated radar measurements." and "5 Experiments We conduct experiments using our real-world ground-based camera dataset, consisting of 17 hours of camera data capturing shallow cumulus clouds across 12 days. |
| Researcher Affiliation | Academia | Jacob Lin Department of Computer Science University of Oxford EMAIL Edward Gryspeerdt Department of Physics Imperial College London EMAIL Ronald Clark Department of Computer Science University of Oxford EMAIL |
| Pseudocode | No | The paper describes its methodology in Section 3 and its subsections (3.1, 3.2, 3.3), but these are presented as descriptive text rather than formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code and data are available on our project page https://cloud4d.jacob-lin.com/." and in the "Open access to data and code" justification: "Open access to our code and datasets will be made prior to the conference. |
| Open Datasets | Yes | Code and data are available on our project page https://cloud4d.jacob-lin.com/." and in the "Open access to data and code" justification: "Open access to our code and datasets will be made prior to the conference. The radar and satellite data used are publicly available." It also mentions using "ERA5 (Hersbach et al., 2020) data" and satellite data. |
| Dataset Splits | No | To train our model, we require ground-based images that are paired with 3D grids representing the cloud liquid water content." (referring to synthetic data) and "To evaluate our method, we collect real-world images from six cameras across a two-month period." The paper describes using synthetic data for training and real-world data for evaluation, but it does not specify explicit train/validation/test splits of these datasets with percentages or sample counts for reproduction, other than the length of the real-world evaluation dataset (17 hours). |
| Hardware Specification | Yes | Optimization is done using Adam (Kingma and Ba, 2015), and takes three days with 4x H100 80GB GPUs. |
| Software Dependencies | No | Our 2D CNN implementation is based on the architecture of EDM Karras et al. (2022)... For the sparse transformer, we follow the architecture of TRELLIS (Xiang et al., 2024)..." and "specifically Co Tracker3 (Karaev et al., 2024)." and "using Adam (Kingma and Ba, 2015)". The paper mentions specific models and algorithms, but does not provide specific version numbers for underlying general software dependencies like Python or PyTorch. |
| Experiment Setup | Yes | Training is performed for 60k steps in the first stage and 30k steps in the second stage. Optimization is done using Adam (Kingma and Ba, 2015), and takes three days with 4x H100 80GB GPUs." and "λCBH and λ h are hyperparameters to scale the losses to similar ranges and are both set to 0.1." and from Appendix A, Table 2: "Learning rate", "Schedule", "Steps", "Optimizer: Adam with (β1, β2) = (0.9, 0.999) Batch Size: 1 Gradient Clipping: 1 Weight Decay: 0 Augmentations" |