以非監督式機器學習模型預測二聚體分子之總能量曲線

Abstract

近幾年間，隨著電腦設備的加強與原子結構逐漸複雜，我們實驗室慢慢的從人為用量子化學的計算方式，結合機器學習，更快的得到未知分子二聚體相互作用能的各種資料。先前實驗室是建立了8種常見的官能基分類數據庫，使用同源二聚體(Dimers)的穩定點資料來預測異源二聚體(Bimers)的穩定點資料，都有得到不錯的結果，但是只有穩定點的資料對於分子動力學的接續來說是不夠的，所以我試著使用機器學習的方式，來預測出每個二聚體之間的能量曲線。 我選擇的是一種前饋式的人工神經網路(ANN)所組成的非監督式機器學習模型，由於是非監督試的機器模型，他並不知道能量曲線的走向，所以需要方法逐步引導機器走向正確的地方，我的工作就是找出能讓機器正確學習的步驟，盡可能地減少訓練集中的參數，能讓機器把正確的二聚體位能曲線給表現出來。 目前只用了實驗室裡的數據集，31個同源二聚體和268個異源二聚體，來做訓練與預測，希望能先找到能正確預測出能量曲線的方法後再去預測其他更大的資料庫，最終目標希望能夠預測出各種分子構型的能量曲線。 最後結果可以知道，並不用把所有參數都丟進訓練集，只需要6或7個單點位置，機器就能夠知道整體曲線的趨勢，不僅使用更少的資料來訓練，預測出的結果也比全部資料拿去訓練的結果更好，所以有次序的進行訓練，是有助於電腦進行學習的。

In the past few years, with the enhancement of computer equipment and the increasing complexity of atomic structures, our laboratory has slowly moved away from artificial quantum chemical calculations and combined with machine learning to obtain various data on the interaction energy of unknown molecular dimers more quickly. Previously, we built a database of 8 common functional groups and used the stability point data of homo-dimers to predict the stability point data of hetero-dimers, and we got good results. Therefore, I would like to use machine learning to predict the energy curve of each dimer. I chose a unsupervised machine learning model composed of a feed-forward artificial neural network (ANN). Since it is a non-supervised machine model, it doesn't know the direction of the energy curve, so it needs a way to gradually guide the machine to the right place, and my job is to find out the steps that allow the machine to learn correctly, and to minimize the number of parameters in the training set as much as possible, and to allow the machine to represent the correct dimer potential energy curve. At present, we only use the laboratory data set, 31 homodimers and 268 heterodimers, for training and prediction. We hope to find a way to predict the energy curves correctly before predicting other larger databases, with the ultimate goal of predicting the energy curves of various molecular configurations. In the final result, we can see that we don't need to throw all the parameters into the training set, only 6 or 7 single point positions are needed, the machine can know the trend of the overall curve, not only using less data for training, but also predicting the result is better than the result of taking all the data to the training, so the sequential training is helpful for the computer to learn.