深層生成模型輔助鋰電池電解質之材料結構設計與性質預測

Abstract

新穎材料的開發能帶給人們許多好處與生活上的便利。傳統上尋找所需化學特性的新穎材料分子從開發至上市約須花費數年甚至數十年。為加速新穎材料開發，計算材料專家引入基於機器學習的反向設計方法，篩選目標材料有助於減少開發成本。本研究開發了一種遞迴神經網路(Recurrent neural networks, RNN)的化學變分自編碼模型(Chemical variational autoencoder, CVAE)，並應用於研究鋰電池內的電解質材料的開發流程，了解不同物理性質與化學結構間之關聯性，我們從QM9分子資料庫篩選出符合特定性質之分子擔任訓練資料集，藉由模型訓練得到一個可生成化學結構與預測材料化學性質之深度學習模型，模型同時可確認其生成化學分子結構的有效性，使化學結構與化學特性彼此建立關聯性，讓模型具有對應的物理意義。對於生成分子的條件上，我們選擇最高佔據分子軌域能階、最低未佔據分子軌域能階、與其能隙等化學性質作為評估生成分子的化學穩定性的重要參數，介電系數參數作亦為抑制電解質內樹枝狀結構生成，而提升化學穩定性的重要參數。結果顯示，模型除了可生成符合特性之化學結構之外，並且藉由模型損失函數之權重的調整，我們的模型具有更準確與更有效率的生成化學結構效能。

The discovery of novel materials brings enormous benefits and convenience to our life. To accelerate the exploration of novel materials, the deep-learning-based inverse design for the intelligent discovery of organic molecules was introduced by experts in computational materials. In our research, a chemical variational autoencoder (CVAE) designed by recurrent neural networks (RNN) was developed and applied to the development process of electrolyte materials in lithium-ion batteries. We screened out molecules with specific properties from the QM9 molecular database as the training data set. Through model training, we obtained a deep learning model that can generate chemical structures with high validity and predict the chemical properties. Besides, the model set up the relation between the chemical structures and desired chemical properties. Regarding the conditions for generating molecules, we selected important chemical properties as important parameters for evaluating the chemical stability for generating molecules purposefully, such as HOMO, LUMO, and HOMO-LUMO gap. Another parameter for chemical stability is the dielectric constant, which is for suppressing the formation of dendritic growth and improving chemical stability. The results showed that in addition to generating chemical structures that match the chemical properties, our model possessed more accurate and efficient ability on the generation performance chemical structures by adjusting the weight of the loss function of deep learning model.