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 [1].

GotenNet: Rethinking Efficient 3D Equivariant Graph Neural Networks

Authors: Sarp Aykent, Tian Xia

ICLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluated models on QM9, r MD17, MD22, and Molecule3D datasets, where the proposed model consistently outperforms state-of-the-art methods in both scalar and high-degree property predictions, demonstrating exceptional robustness across diverse datasets, and establishes Goten Net as a versatile and scalable framework for 3D equivariant Graph Neural Networks.
Researcher Affiliation Industry Sarp Aykent1 and Tian Xia2 1Comcast AI Technologies, 2Microsoft
Pseudocode No The paper includes architectural diagrams (Figure 2) and mathematical equations to describe the methodology, but it does not contain a clearly labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code Yes https://github.com/sarpaykent/Goten Net
Open Datasets Yes We evaluated models on QM9, r MD17, MD22, and Molecule3D datasets... The proposed method is evaluated against a comprehensive set of baselines using the QM9 dataset (Ruddigkeit et al., 2012; Ramakrishnan et al., 2014)... We further evaluate our model on the Molecule3D dataset (Xu et al., 2021)... The MD22 dataset (Chmiela et al., 2023) contains molecular dynamics trajectories... The r MD17 dataset (Christensen & Von Lilienfeld, 2020) is a revised version of the MD17 benchmark
Dataset Splits Yes The r MD17 dataset (Christensen & Von Lilienfeld, 2020) is a revised version of the MD17 benchmark, featuring 10 small organic molecules with 100,000 conformations per molecule. It serves as a key benchmark for evaluating machine learning models ability to predict molecular energies and forces across diverse conformations. We follow the standard split (Christensen & Von Lilienfeld, 2020) of 950 training, 50 validation, and the remaining conformations for testing. The results are averaged over five predefined splits to ensure robust evaluation.
Hardware Specification Yes Experiments were conducted with an NVIDIA A100 GPU with 80GB video memory, 512GB RAM, and an AMD EPYC 7713P CPU.
Software Dependencies No The paper mentions "spherical harmonics were computed using the e3nn (Geiger & Smidt, 2022) library," but it does not specify a version number for the e3nn library or any other software dependencies.
Experiment Setup Yes Table 6: Hyper-parameters for the datasets Goten Net compared against the baselines. The parameters are for Goten Net B if multiple variations exists. Hyper-parameters: Optimizer (Adam W), Learning rate scheduling (Linear warmup with reduce on plateau), Warmup steps (10,000 to 1,000), Maximum learning rate ([6e-5, 1e-4] to 2e-4), Learning rate decay (0.8), Learning rate patience (15 to 30), Loss function (MSE, L1), Gradient clipping (10, 5), Batch size (32 to 256), Number of epochs (1,000 to 3,000), Weight decay (0.01), Dropout rate (0.1), Node dimension (192 to 384), Edge dimension (192 to 384), Edge refinement dimension (256 to 768), Lmax (2), Number of Layers (6 to 12), Number of RBFs (32 to 64), Number of Attention Heads (8), Cutoff radius (4.0 to 5.0).