應用分子的sigma profile建立機器學習模型於預測二氧化碳在深共熔溶劑中之溶解度;Prediction of CO2 Solubility in Deep Eutectic Solvents from Machine Learning Models with Molecular σ-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 987654321/97516

請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/97516

題名:	應用分子的sigma profile建立機器學習模型於預測二氧化碳在深共熔溶劑中之溶解度;Prediction of CO2 Solubility in Deep Eutectic Solvents from Machine Learning Models with Molecular σ-profiles
作者:	陳穎蓉;Chen, Ying-Jung
貢獻者:	化學工程與材料工程學系
關鍵詞:	機器學習;二氧化碳捕捉;深共熔溶劑;COSMO-SAC 模型;machine learning;CO2 capture;deep eutectic solvents;COSMO-SAC model
日期:	2025-07-26
上傳時間:	2025-10-17 11:28:25 (UTC+8)
出版者:	國立中央大學
摘要:	碳捕捉作為一項重要的環境治理措施和應對氣候變遷的關鍵策略，成為目前研究的重視項目。工業中常用的醇胺溶液在碳捕捉領域已經具有成熟的應用基礎，但因為高腐蝕性、高能耗以及溶劑損耗等問題限制了其在未來的發展。這些技術在大規模應用中往往需要投入大量的能源，對實現低碳社會形成一大挑戰。因此深共熔溶劑具備低毒性、低揮發性、可調控性及良好的化學穩定性等特點，逐漸被認為是代替醇胺溶液的潛在候選者，尤其是在捕捉二氧化碳方面展現出良好的應用前景。然而深共熔溶劑的組成多樣性對於研究二氧化碳溶解度及性能優化提出了更高要求。為了解決這一挑戰，本研究採用機器學習技術結合量子化學計算，對二氧化碳在不同組成中的深共熔溶劑進行溶解度預測。研究中，我們測試了不同的機器學習模型以及兩種資料切分方式，用於分析和預測二氧化碳在深共熔溶劑中的溶解行為。為了進一步提高模型的準確性與穩定性，我們應用了源自量子化學計算的分子描述符，基於COSMO-SAC 2010模型計算獲得分子表面電荷分布做為分子的描述符，產生的分子描述符能夠有效捕捉深共熔溶劑的結構特徵。為了驗證模型性能，我們建立了一個包含2,520筆數據的資料庫，該資料庫涵蓋了127種深共熔溶劑在不同溫度、壓力範圍及組成條件下的二氧化碳溶解度數據。收集的資料庫分布廣泛，充分涵蓋了實際應用中典型的化學組成和操作條件。藉由將模型的預測結果與實驗數據進行對比，我們發現結合量子化學計算與機器學習的方法，不僅能顯著提升二氧化碳溶解度的預測準確性，還能有效解決傳統方法在處理複雜系統的局限性。在所有模型中，XGBoost表現最佳，在資料樣本切分下，R²達到0.984，RMSE為0.078；在更具挑戰性的溶劑組成切分下，也有良好表現，R²為0.924，RMSE為0.168。預測結果充分地說明分子描述符的選擇對模型性能具有明顯的影響，再次證明量子化學計算能夠為二氧化碳捕捉劑的篩選和設計提供分子層面資訊。;Carbon capture is a key strategy for mitigating climate change and promoting sustainable environmental management. While amine-based solvents are widely used in industrial CO₂ capture, issues such as high energy demand, corrosiveness, and solvent degradation hinder their long-term viability. Deep eutectic solvents have emerged as promising alternatives due to their low toxicity, low volatility, chemical stability, and tunability. However, the chemical complexity and structural diversity of DESs present challenges in accurately predicting CO₂ solubility and optimizing solvent performance. To address these challenges, this study integrates machine learning techniques with quantum chemical calculations to develop predictive models for CO₂ solubility in DESs. Different machine learning models were tested with two data splitting strategies, and molecular descriptors derived from COSMO-SAC 2010 model were employed. Sigma profiles representing surface charge density distributions were used to capture molecular characteristics of DESs. A comprehensive dataset was compiled containing 2,520 CO₂ solubility records across 127 unique DESs under various temperatures, pressures, and molar ratios. This dataset ensured broad chemical diversity and coverage of practical operating conditions. Model performance was evaluated by comparing predictions with experimental data. The results showed that integrating quantum chemical descriptors with machine learning significantly improved predictive accuracy and generalization. Among all the models evaluated, XGBoost achieved the best performance. Under the sample-wise splitting strategy, the model attained an R2 of 0.984 with an RMSE of 0.078. When applying the more challenging DES-wise splitting strategy, the performance remained strong, with an R2 of 0.924 and an RMSE of 0.168. This study highlights the importance of descriptor selection and demonstrates that quantum chemical calculations provide meaningful molecular-level insights for designing and screening effective CO₂ capture solvents. The proposed framework offers a scalable and accurate approach to support the development of next-generation DES-wise carbon capture systems.
顯示於類別:	[化學工程與材料工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	23	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....