建立量子化學交互作用能資料集應用於機器學習勢能

Abstract

準確的非共價相互作用對於超分子化學、材料科學和生物化學等領域皆是研究的重點，而過去實驗室也針對不同的二聚體分子進行大量的相互作用能計算，並建立了SOFG-31資料集，而為了應對廣泛的分子系統，我們需要建構更大的資料集。因此在本研究第一部分，我們建立了DES133資料集，包含了39220筆非共價相互作用能與其結合能量（靜電、交換、誘導和色散），其中涵蓋1961個獨特的二聚體，並使用高階對稱適應微擾理論SAPT銅(SAPT2/jVTZ)與銀(SAPT2+/aVDZ)標準進行計算，最後根據SAPT的計算結果進行能量分析。此外，基於先前的機器學習力場技術(AP-Net)的開發，並在先前的研究中透過特定的方式以數十筆的資料來預測數百筆的資料。因此，在研究的第二部分，我們使用此方法透過DES133資料集中數百筆同源二聚體分子作為訓練集，以預測相同資料集中的數千筆異源二聚體。結果顯示，整體的預測誤差低於化學精度1kcal/mol，僅用特定二聚體分子作為訓練就可重現各種二聚體相互作用能，因此透過此種方法，我們就能以更少的時間與計算成本建構一個機器學習力場，以應用於各種分子間勢能表面。

Accurate non-covalent interactions are crucial for research in fields such as supramolecular chemistry, materials science, and biochemistry. Our laboratory has previously conducted extensive interaction energy calculations for various dimer molecules, resulting in the establishment of the SOFG-31 datasets. However, to address a wider range of molecular systems, a larger dataset is necessary. Therefore, in the first part of this study, we have developed the DES133 dataset, which includes 39220 noncovalent interaction energies and their binding energy components (electrostatic, exchange, induction, and dispersion). This dataset includes 1961 unique dimers and was calculated using the higher-order symmetry-adapted perturbation theory (SAPT) bronze (SAPT2/jVTZ) and silver (SAPT2+/aVDZ) standards, followed by energy analysis based on the SAPT results. Furthermore, based on the preceding development of Machine Learning Force Field Technology (AP-Net) and the preceding study, we predicted a multitude of data points, ranging from hundreds to tens, in a specific manner. Consequently, in the second part of the study, this method was employed to predict thousands of heterodimers in the same dataset, utilising hundreds of homodimers in the DES133 dataset as a training set. The results demonstrate that the overall prediction error is less than the chemical accuracy of 1 kcal/mol, and a range of dimer interaction energies can be reproduced by utilising a single specific dimer molecule as the training set. Consequently, this approach enables the construction of a machine-learning force field for diverse intermolecular potential energy surfaces in a more time-efficient and cost-effective manner.