arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2104.10869cond-mat.mtrl-sciphysics.comp-ph
收藏

CrysXPP:一个可解释的晶体材料性质预测器

CrysXPP:An Explainable Property Predictor for Crystalline Materials

Kishalay Das, Bidisha Samanta, Pawan Goyal, Seung-Cheol Lee, Satadeep Bhattacharjee, Niloy Ganguly

摘要:

我们提出了一个深度学习框架CrysXPP,允许以合理的精度快速预测各种材料的电子,磁性和弹性特性。尽管我们的工作与最近几次构建基于深度学习的属性预测器的尝试是一致的,但它克服了它们的一些局限性。CrysXPP通过智能设计自动编码器CrysAE并将结构信息传递到属性预测过程,降低了对大量标记数据的需求,从而训练深度学习模型。自动编码器又在大量未标记的晶体图上进行训练,设计的损失函数有助于捕获所有重要的结构和化学信息。此外,CrysXPP仅使用少量标记数据进行属性预测,还训练了一个特征选择器,为获得的结果提供可解释性。我们证明,CrysXPP令人信服地表现优于所有竞争和最近的基线算法在七个不同的属性集。最值得注意的是,当给出少量实验数据时,CrysXPP始终能够优于传统DFT。我们发布了大型预训练模型CrysAE,以便研究社区可以在具有受限数据源的各种应用程序上使用少量标记数据对其进行微调。

英文摘要:

We present a deep-learning framework, CrysXPP, to allow rapid prediction of electronic, magnetic and elastic properties of a wide range of materials with reasonable precision. Although our work is consistent with several recent attempts to build deep learning-based property predictors, it overcomes some of their limitations. CrysXPP lowers the need for a large volume of tagged data to train a deep learning model by intelligently designing an autoencoder CrysAE and passing the structural information to the property prediction process. The autoencoder in turn is trained on a huge volume of untagged crystal graphs, the designed loss function helps in capturing all their important structural and chemical information. Moreover, CrysXPP uses only a small amount of tagged data for property prediction, and also trains a feature selector that provides interpretability to the results obtained. We demonstrate that CrysXPP convincingly performs better than all the competing and recent baseline algorithms across seven diverse set of properties. Most notably, when given a small amount of experimental data, CrysXPP is consistently able to outperform conventional DFT. We release the large pretrained model CrysAE so that it could be fine-tuned using small amount of tagged data by the research community on various applications with restricted data source.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑