利用SAPT 資料集建立基於機器學習與物理模型的分子間交互作用能計算方法

Abstract

先前透過量子化學計算方法已建立常見官能基二聚體交互作用能之數據庫，但隨著分子結構的增大，基於此方法的計算成本倍數增加，因此實驗室的數據庫基本停留在十數個到二十個重原子左右。我們不希望止步與此，但是繼續使用CPU-based的量子化學計算方法又有所困難，因此我們嘗試以現有的數據庫結合機器學習的方法，以得到計算成本低且穩定可靠的替代方法。 我們選擇了結合機器學習方法以及交互作用能物理模型相結合的計算方法，相較於純粹的機器學習，此法更能控制結果並觀察相關的物理意義。我們利用量子化學計算中的對稱性匹配微擾理論(SAPT)，其將交互作用拆分為四個不同物理意義的作用，分別為是靜電能、交換能、色散能與誘導能，利用數學式描述這四項交互作用的物理現象，並以SAPT所計算的能量對物理模型擬合，以此得到能夠更快計算分子間交互作用能的手段。 我們相信，實驗室過去以常見官能基、二聚體的總原子數由小到大等規則所建立的數據庫應能更好的訓練模型，並且為了應付更大範圍的分子種類，我們自其他學者所公布的資料集中挑選相似性質的資料進行補充，並將結果測試在各常見之資料集，以此驗證結果的可靠度。

Previously, a database of common functional group dimer interaction energies was established through quantum chemical calculation methods, but with the increase of the size of molecular structure, the computing cost based on this method become higher and higher. Therefore, our database stayed between 10-20 heavy atoms’ dimers. This is not what we want to see, but it is difficult to keep using the CPU-based quantum chemical calculation methods. So we try to combine our current database with machine learning, hoping to get a reliable and stable alternative method. We choose a computational approach that combines a machine learning method with some physics models of the interaction energy, which allows more control over the results and observation of the relevant physical meanings. We use the Symmetry-Adapted Perturbation Theory(SAPT) in quantum chemical calculations, which divides the interaction into four terms with different physical meanings, namely electrostatic energy, exchange energy, dispersion energy, and induced energy. By using the mathematical formula described those four interactions’ physical phenomena, supplemented by fitting the model with the value calculated from SAPT. Then we get a method that can compute higher molecular weight dimers at a low computational cost. We believe that the database established by the laboratory in the past which complies with the rules of common functional groups and the total number of atoms of dimer molecules that grow from small to large should be able to adequately train the model. And to cope with a wider range of molecular species, we use some similar properties’ database to supplement. Then results are tested in various common datasets to verify the reliability.