Retrieval-Retro: Retrieval-based Inorganic Retrosynthesis with Expert Knowledge
Authors: Heewoong Noh, Namkyeong Lee, Gyoung S. Na, Chanyoung Park
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate the superiority of Retrieval-Retro in retrosynthesis planning, especially in discovering novel synthesis recipes, which is crucial for materials discovery. |
| Researcher Affiliation | Academia | Heewoong Noh1, Namkyeong Lee1, Gyoung S. Na2 , Chanyoung Park1 1 KAIST 2 KRICT |
| Pseudocode | Yes | Algorithm 1: Pseudocode of Retrieval-Retro. Algorithm 2: Pseudocode of MPC Retriever. Algorithm 3: Pseudocode of NRE Retriever. |
| Open Source Code | Yes | The source code for Retrieval-Retro is available at https://github.com/Heewoong Noh/Retrieval-Retro. |
| Open Datasets | Yes | Datasets. We use 33,343 inorganic material synthesis recipes extracted from 24,304 materials science papers [20] following prior studies [12, 18]. For DFT-calculated data, we use Materials Project [14] database 3, which is an openly accessible database that provides various material properties calculated using DFT. |
| Dataset Splits | Yes | Following prior studies, under the random split setting, we randomly split the dataset into train/valid/test of 80/10/10%. On the other hand, under the year split setting [12], the training set includes synthesis recipes from papers published up to 2014, the validation set includes recipes from papers published in 2015 and 2016, and the test set includes recipes from papers published between 2017 and 2020. |
| Hardware Specification | Yes | All experiments are conducted on a 48 GB NVIDIA RTX A6000. |
| Software Dependencies | Yes | Our method is implemented on Python 3.8.13, and Torch-geometric 2.0.4. |
| Experiment Setup | Yes | We train the model for 500 epochs across all tasks, while the model is early stopped if there is no improvement in the best validation Top-5 Accuracy for 30 consecutive epochs. Table 5: Hyperparameter specifications of Retrieval-Retro. |