開發基於分子Sigma Profile的深度神經網絡以估算PC-SAFT狀態方程式的純物質參數;Estimating Pure Component Parameters of PC-SAFT EOS from Deep Neural Network with Molecular Sigma Profiles

NCU Institutional Repository > 工學院 > 化學工程與材料工程研究所 > 博碩士論文 > Item 987654321/97524

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/97524

题名:	開發基於分子Sigma Profile的深度神經網絡以估算PC-SAFT狀態方程式的純物質參數;Estimating Pure Component Parameters of PC-SAFT EOS from Deep Neural Network with Molecular Sigma Profiles
作者:	陳家豪;Chen, Jia-Hao
贡献者:	化學工程與材料工程學系
关键词:	熱力學;擾動鏈統計關聯流體理論狀態方程式;分子特定表面電荷密度分布;機器學習;深度神經網路;Thermodynamics;PC-SAFT EOS;Sigma Profile;Machine Learning;Deep Neural Network
日期:	2025-07-28
上传时间:	2025-10-17 11:28:51 (UTC+8)
出版者:	國立中央大學
摘要:	熱力學模型在預測材料性質方面具有顯著的優勢，能在不需大量實驗的情況下提供可靠結果，不僅節省時間，亦可降低經濟成本，因此是一種高效的材料性質預測方法。在本研究中，我們選擇了PC-SAFT (擾動鏈統計關聯流體理論)模型來預測物質性質，該模型在預測複雜流體(特別是含有長鏈碳氫化合物及高分子)的行為方面表現優異。此外，PC-SAFT所需的經驗參數通常較少，且這些參數具有明確的物理意義。接著，我們利用深度神經網路來估算PC-SAFT方程式的純物質參數。採用此方法的原因是現有文獻中可取得的PC-SAFT純物質參數數量有限，尤其對於某些複雜化合物而言，缺乏實驗數據。深度學習使我們即便僅有少量實驗數據，仍能獲得良好的預測結果，並且可進一步預測這些複雜化合物的參數。本研究結合量子化學COSMO溶合計算與深度神經網路(DNN)，建立一套以分子Sigma Profile為描述子的PC-SAFT狀態方程式純物質參數預測方法，純物質參數包含三個非關聯參數及兩個關聯參數。研究首先收集並擴增包含非締合與締合分子的數據庫，取得分子Sigma Profile、表面積與體積等特徵作為模型輸入，針對片段數、片段直徑、分散能量及關聯參數分別建立預測模型。過程中透過不同訓練集比例以及不同數據集的調整進行分組比較，並與文獻官能基貢獻法及分子指紋方法進行誤差指標比較，同時分析參數分佈特性與模型穩健性。此外，研究進一步以預測結果誤差調整參數進行相平衡計算，驗證飽和蒸氣壓與液體密度之應用可行性。整體結果顯示，本方法可有效應用於非締合分子PC-SAFT參數預測，並為後續熱力學模型資料庫擴充與新物質開發提供支持。而部分複雜分子或關聯參數的預測仍受限於數據分佈與數據量，期望未來可以透過擴增數據等方式進一步改善。 ;Thermodynamic models offer significant advantages in predicting material properties, providing reliable results without the need for extensive experimental work. This not only saves time but also reduces costs, making them an efficient tool for material property prediction. In this study, we selected the PC-SAFT (Perturbed Chain-Statistical Associating Fluid Theory) model to predict substance properties, as it performs well in describing the behavior of complex fluids, especially those containing long-chain hydrocarbons and polymers. Moreover, the empirical parameters required by PC-SAFT are usually fewer in number and have clear physical meaning. To estimate these parameters, we applied deep neural networks, since the number of available PC-SAFT pure component parameters reported in the literature is limited, particularly for certain complex compounds that lack experimental data. The use of deep learning enables satisfactory prediction results even with a small amount of experimental data and allows the estimation of parameters for these complex substances. This research integrates quantum chemical COSMO solvation calculations with deep neural networks (DNNs) to develop a prediction approach for PC-SAFT pure component parameters based on the molecular Sigma Profile as a descriptor. The parameters include three non-associating and two associating parameters. We first collected and expanded a database covering both non-associating and associating compounds, obtained Sigma Profiles, surface areas, and volumes as input features, and then constructed prediction models for the segment number, segment diameter, dispersion energy, and associating parameters. The models were trained and compared under different training set sizes and dataset configurations, and benchmarked against group contribution methods and molecular fingerprint methods reported in the literature. In addition, parameter distribution characteristics and model robustness were analyzed. Furthermore, the predicted results were adjusted for errors and applied to vapor–liquid equilibrium calculations to verify the feasibility of predicting vapor pressure and liquid density. Overall, this method demonstrates the capability to effectively predict PC-SAFT parameters for non-associating compounds and provides support for expanding thermodynamic model databases and the development of new substances. The prediction of certain complex molecules or associating parameters is still limited by data distribution and sample size, which is expected to be further improved through future data expansion.
显示于类别:	[化學工程與材料工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	19	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....