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..
A Physics-preserved Transfer Learning Method for Differential Equations
Authors: Hao-Ran Yang, Chuan-Xian Ren
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments in simulation and real-world datasets demonstrate the superior performance, generalizability and physics preservation of the proposed POTT method. 5 Experiment Benchmarks. Experiments are conducted on both simulation and real-world datasets. For simulation datasets, three representative equations are contained. For real-world datasets, we consider the cross-region climate forecasting task. |
| Researcher Affiliation | Academia | Hao-Ran Yang Sun Yat-Sen University Guangzhou, China EMAIL Chuan-Xian Ren Sun Yat-Sen University Guangzhou, China EMAIL |
| Pseudocode | Yes | B.3 Algorithm analysis Algorithm 1: Optimization of POTT 1 Input: source data ˆDs, pretrained model Gη, target data ˆDt; 2 Initialize Tθ, fϕ; 3 for N11 steps do 4 Freeze ϕ, update θ to minimize the first objective of Eq. (14) for N12 steps; 5 Freeze θ, update ϕ to maximize the first objective of Eq. (14) with θ; 7 for N2 steps do 8 Freeze θ and ϕ, update η to minimize the second objective in Eq. (14); 10 Output: Gη as approximation of Gt. |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: As our work is about a general model transfer method for DEs problems, the reproducibility can be guaranteed by the algorithm description and implementation details provided in Sec. 5, Appendix B.3 and Appendix C. |
| Open Datasets | Yes | Simulation dataset. Following previous works [22, 25], we take the 1-d Burgers equation, 1-d space-time Advection equation, and 2-d Darcy Flow problem as our benchmarks. Real-world dataset. ... The preprocessed 5.625 resolution ERA5 dataset from Weather Bench [29] is used for evaluation and the SOTA Clim ODE [33] is used as backbone model. |
| Dataset Splits | Yes | We generate 1000 training samples for each domain of Burgers equation and 2000 samples for Advection equation and Darcy Flow. To fully investigate the effectiveness of transfer learning methods, we consider two scenarios that only 50 and 100 target data samples are available for model transfer. For all transfer tasks, we use 10 extra target domain samples for validation and 100 for testing. For source domain, we use ten years of training data (2006-15). For target domain, only one year (2015) of training data is available, while data of 2016 is used for validation and data of 2017-18 is used as testing data. |
| Hardware Specification | Yes | All experiments are conducted on a single 16GB NVIDIA 4080 device. |
| Software Dependencies | No | We use Adam as optimizer and the learning rate is 1e-3 for all tasks. The learning rate of the backbone of Gη is ten times smaller than the last two layers, which is a widely-used technique in transfer learning. A cosine annealing strategy is adopted for learning rate of Gη. Details of model architectures and data generation can be found in codes provided by [22, 25]. |
| Experiment Setup | Yes | We use Adam as optimizer and the learning rate is 1e-3 for all tasks. The learning rate of the backbone of Gη is ten times smaller than the last two layers, which is a widely-used technique in transfer learning. A cosine annealing strategy is adopted for learning rate of Gη. |