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..
Learning 3D Anisotropic Noise Distributions Improves Molecular Force Fields
Authors: Xixian Liu, Rui Jiao, ZHIYUAN LIU, Yurou Liu, Yang Liu, Ziheng Lu, Wenbing Huang, Yang Zhang, Yixin Cao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that Ani DS outperforms prior isotropic and homoscedastic denoising models and other leading methods on the MD17 and OC22 benchmarks, achieving average relative improvements of 8.9% and 6.2% in force prediction accuracy. |
| Researcher Affiliation | Collaboration | 1Fudan University, 2Tsinghua University, 3National University of Singapore, 4Renmin University of China, 5 Microsoft Research 6Institute of Trustworthy Embodied AI, Fudan University |
| Pseudocode | No | The paper describes methodology and theoretical derivations with equations and diagrams (e.g., Figure 2), but does not contain explicitly labeled "Pseudocode" or "Algorithm" blocks with structured steps. |
| Open Source Code | Yes | Our code is available at https://github.com/Zero Knighting/Ani DS. |
| Open Datasets | Yes | Our experiments involve four datasets: (1) PCQM4Mv2 [32] contains 3,746,619 molecules along with their Density Functional Theory (DFT) calculated 3D equilibrium structures. (2) OC22 [33] is a large-scale dataset of DFT calculated structures designed to advance machine learning for oxide electrocatalysts. (3) MD17 [34] provides molecular dynamics trajectories of small organic molecules, with both energy and force labels. (4) MPtrj [35] includes 1.58 million structures obtained from DFT relaxation trajectories of over 146,000 materials in the Materials Project [36]. |
| Dataset Splits | Yes | Following [27], we use 950 molecules for training and 50 for testing. No noise is added during validation and testing. ... Additionally, our results on MPTrj are provided in Appendix C. |
| Hardware Specification | Yes | Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: These details are provided in Appendix C.2. |
| Software Dependencies | No | The main text of the paper does not provide specific software names with version numbers or other detailed software dependencies. The NeurIPS checklist indicates details are in Appendix C.2, but without access to the appendix, this information is not explicitly provided in the core paper content. |
| Experiment Setup | Yes | Following [27], we use 950 molecules for training and 50 for testing. No noise is added during validation and testing. ... with the best performance achieved at ω = 0.1. In addition, Table 3c (bottom) shows the effect of different ϖKL values. Our results indicate that setting ϖKL = 1.0 yields the best performance. |