Crystalformer: Infinitely Connected Attention for Periodic Structure Encoding

Authors: Tatsunori Taniai, Ryo Igarashi, Yuta Suzuki, Naoya Chiba, Kotaro Saito, Yoshitaka Ushiku, Kanta Ono

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Quantitative comparisons using Materials Project and JARVIS-DFT datasets show that the proposed method outperforms several neural-network-based state-of-the-art methods (Xie & Grossman, 2018; Schütt et al., 2018; Chen et al., 2019; Louis et al., 2020; Chen & Ong, 2022; Choudhary & De Cost, 2021; Yan et al., 2022) for various crystal property prediction tasks.
Researcher Affiliation Collaboration Tatsunori Taniai1, Ryo Igarashi1, Yuta Suzuki2, Naoya Chiba3, Kotaro Saito4,5 Yoshitaka Ushiku1, Kanta Ono5 1OMRON SINIC X Corp., 2Toyota Motor Corp., 3Tohoku University, 4Randeft Inc., 5Osaka University
Pseudocode No The paper describes methods and architectures but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes We release our code online.
Open Datasets Yes Materials Project (MEGNet) is a collection of 69,239 materials from the Materials Project database retrieved by Chen et al. (2019)." and "JARVIS-DFT (3D 2021) is a collection of 55,723 materials by Choudhary et al. (2020).
Dataset Splits Yes Thanks to the great effort by Choudhary & De Cost (2021) and Yan et al. (2022), many relevant methods are evaluated on these datasets with consistent and reproducible train/validation/test splits." and "The three numbers below each property name indicate the sizes of training, validation, and testing subsets.
Hardware Specification Yes We evaluated the running times of all the methods by using a single NVIDIA RTX A6000 and a single CPU core on the same machine.
Software Dependencies No The paper mentions optimizers and activation functions but does not specify any software library names with version numbers required for reproduction.
Experiment Setup Yes For each regression task in the Materials Project dataset, we train our model by optimizing the mean absolute error loss function via stochastic gradient descent (SGD) with a batch size of 128 materials for 500 epochs. We initialize the attention layers by following Huang et al. (2020). We use the Adam optimizer (Kingma & Ba, 2015) with weight decay (Loshchilov & Hutter, 2019) of 10-5 and clip the gradient norm at 1. For the hyperparameters of Adam we follow Huang et al. (2020); we use the initial learning rate α of 5 × 10-4 and decay it according to α √4000/(4000 + t) by the number of total training steps t, and use (β1, β2) = (0.9, 0.98).