| 摘要: | 本研究針對PC-SAFT狀態方程式純物質參數進行預測，PC-SAFT又分為有氫鍵系統與無氫鍵系統，有氫鍵系統需要五個純物質參數m、σ、ϵ、 ϵAB與кAB，無氫鍵系統則需要三個純物質參數m、σ、ϵ，本文只探討PC-SAFT無氫鍵系統的部份。我們對兩種機器學習方法進行討論與比較，分別為深度神經網絡(deep neural network, DNN)與隨機森林(random forest, RF)。我們提出機器學習與熱力學模型COSMO溶合計算結合，透過COSMO計算後產生的Sigma Profile作為機器學習的特徵進而對純物質參數進行預測。
本研究收集了1153個無氫鍵鍵結的分子Sigma Profile與其PC-SAFT的三個純物質參數，作為機器學習的數據集。我們將訓練模型的大小分成兩組，將數據集以70%作為訓練集30%測試集為第一組，80%作為訓練集20%測試集作為第二組，共兩組。在以分子特徵作為輸入除了Sigma Profile外，加上透過COSMO計算得到的面積與體積(Sigma Profile + area + volume)。在深度神經網絡與隨機森林超參數的優化，我們使用網格搜索(grid search, GS)與貝葉斯優化(bayesian optimization, Byn) 進行模型調整參數並做比較。在70%的訓練的系統下表現最好的為GS搭配DNN:GS+DNN-m的誤差(AARD%)為8.61%，GS+DNN-σ與GS+DNN- ϵ的誤差為3.23%與6.41%。而在80%的訓練系統下，整體誤差最低的也是GS搭配DNN:GS+DNN-m的誤差為8.68%，GS+DNN-σ與GS+DNN- ϵ的誤差為3.06%與6.41%，我們提出的COSMO溶合計算結合機器學習的模型非常穩定，而DNN比RF提供更準確的預測結果。
關鍵字: 機器學習、深度神經網絡、隨機森林、COSMO溶合計算、PC-SAFT狀態方程式、熱力學
;This study focuses on predicting pure component parameters of the PC-SAFT (perturbed chain statistical associating fluid theory) equation of state for the non-hydrogen bonding system. PC-SAFT has two variants, one for systems with hydrogen bonding that requires five pure component parameters (m, σ, ε, ϵAB, and кAB), and another for systems without hydrogen bonding that requires three pure component parameters (m, σ, and ε). This study specifically explores the non-hydrogen bonding system of PC-SAFT. This research discusses and compares two machine learning methods, deep neural network (DNN) and random forest (RF). We propose a combination of machine learning and the thermodynamic model COSMO-SAC, where the Sigma Profile calculated by COSMO-SAC is used as a feature for machine learning to predict the pure component parameters. The dataset for machine learning consists of 1153 Sigma Profiles of molecules without hydrogen bonding and their corresponding three pure component parameters for PC-SAFT. The dataset is divided into two groups for training the models: the first group uses 70% of the data as the training set and 30% as the test set, while the second group uses 80% as the training set and 20% as the test set. For the optimization of hyperparameters in DNN and RF, we employ grid search (GS) and bayesian optimization (Byn) methods to tune the model parameters and make comparisons. Under the 70% training system, the best-performing model is GS combined with DNN (GS+DNN-m) with the an AARD% (average absolute relative deviation) of 8.61%, GS+DNN-σ is 3.23% and for GS+DNN-ε is 6.41%. Under the 80% training system, the overall lowest error is also observed with GS and DNN, for GS+DNN-m is 8.68%, GS+DNN-σ is 3.06% and for GS+DNN-ε is 6.41%. The proposed model, combining COSMO solvation calculation result with machine learning, demonstrates great stability, and DNN provides more accurate predictions compared to RF.
Key words: Machine Learning、Deep Neural Network、Random Forest、COSMO Calculation、PC-SAFT Equation of State、Thermodynamic |
