利用分程深度特徵提取預測分子生成熱

Abstract

增量理論(Increment theory)基於分子中存在的片段來估算其熱力學性質，並且藉由能進行線性加成的特性來預測難以從實驗或理論計算中得出的大分子的熱力學性質，然而增量理論的應用僅限於欲分析的分子有包含預先定義的原子團或成分，如果分子中存在新的化學片段，則需要定義新的原子團並收集新的數據以定義出這些新的化學單元對整體分子性質的貢獻，為了解決這個問題，我們引入了一種機器學習方法，該方法是利用圖像卷積神經網絡（GCNNs）並基於包含原子周圍的局部化學環境來學習每個原子的貢獻，並對每個原子貢獻求和以得出分子的總體性質，由於此方法遵循增量理論的可加成性，因此可以輕鬆地將其推廣到訓練數據中不存在的大分子。除此之外，我們還引入了另一種機器學習架構，該架構可以分別學習局部和全局的分子結構信息使得我們不僅可以得到每個原子的特性，還可以確定整個分子結構對分子性質的貢獻。經過一系列的探討、研究和比較，我們所建構的改良圖像卷積神經網絡在預測熱力學性質的準確度上可以勝過傳統的增量理論。相對於其他現行已開發的熱力學參數預測模型，我們所提出的架構可以提供更精準的預測效能，即使在進行大分子的性質預測或泛化能力檢測時，也可以提供高準確性的熱力學參數評估。

The increment theory estimates the property of a species based on the fragments presented in a molecule and can naturally be scaled up to predict properties of large molecules that are difficult to derive from experiments or ab initio calculations. However, the application of increment theory is limited to the species containing the pre-defined groups or components. If new chemical fragments exist in the molecules, it is required to define new groups and to collect new data to determine the contribution of these new chemical units. To address this issue, we introduce a machine learning approach that can learn the contribution of each atom based on the local environment that encompassing the atom using graph convolutional neural networks (GCNNs), and sum over the atomic contributions to derive the property of the molecule. Because this approach follows the additivity scheme of increment theory, it can be easily generalized to larger molecules that are not present in the training data. In addition, we also introduce a machine learning architecture that can separately learn local and global structural information, and thus can determine not only the properties of each individual atom but also the contribution of the overall molecular structure. We examined the performance of the proposed models on formation enthalpy predictions and found that these models can outperform the conventional increment theory. In addition, they also perform competitively with previously reported machine learning-based thermochemistry estimator, and can achieve a higher accuracy when tested on out-of-domain test sets containing molecules that are significantly larger than those in the training set.