請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/88001
標題: | ML-GA演算法於二氧化碳還原用奈米多孔銅試片製備參數優化之應用 An ML-GA model for applications in optimizing process parameters for nanoporous Cu electrodes in CO2 Reduction |
作者: | 陸俊亦 Jun Yi Lok |
指導教授: | 鄭憶中 I-Chung Cheng |
關鍵字: | 奈米多孔銅(NPC),二氧化碳還原反應(CO2RR),機器學習(ML),基因演算法(GA),參數優化, Nanoporous copper (NPC),Carbon dioxide reduction reaction (CO2RR),Machine learning (ML),Genetic algorithm (GA),Parameter optimization, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 自工業革命以來,燃燒石化產品以及各型工業製造過程所帶來的碳排放已對地球上的自然環境造成了破壞,溫室效應也對人類的福祉帶來了嚴重影響。各個研究團隊也針對降低作為其中一個溫室氣體的二氧化碳含量進行各樣研究,其中二氧化碳還原反應(CO2 Reduction Reaction,CO2RR )也成為了其中一個重點項目。其中,奈米多孔銅(Nanoporous copper,NPC)也因其高反應面積、極佳導電特性、加上銅在CO2RR中能夠產出高經濟價值的碳碳氫化合物(乙烯、乙醇等),在此領域中具有非常大的發展潛能。由於尋找最合適的催化模式需要透過不斷試錯,因此機器學習的加入也能夠降低資源消耗、同時帶來精准預測結果。多數將CO2RR與機器學習相連的研究都偏向最佳催化劑尋找與設計,鮮少有文獻針對CO2RR用催化劑試片製備優化做探討;探討製備過程的優化更能將CO2RR帶入工業上的大型應用。
本研究針對NPC用於CO2RR的基礎上,透過加入機器學習調整與優化NPC試片的製備參數,以便能夠得到最高CO2RR的二碳產物法拉第效率總和的試片。本研究先是針對兩種不同原子比例的銅鋁合金所製備的NPC所完成的CO2RR結果進行整合,並且歸納出了五項試片製備參數作為輸入,以及五種產物的法拉第效率作為輸出以供機器學習模型進行學習。透過利用機器學習(Machine Learning,ML)的預測能力與基因演算法(Genetic Algorithm,GA)的演化性質形成聯合機器學習-基因演算法(ML-GA)模型,使其能夠在指定目標範圍內形成優化後的試片製備參數。在對比三種演算法所形成的ML-GA模型的結果後,由極端梯度提升回歸法(Extreme Gradient Boosting Regression,XGBR)所組成的XGBR-GA演算法所預測的最高二碳產物法拉第效率在不超過物理限制的情況下為三者最高;其總和(乙烯、乙醇與乙酸相加) 預測值為68.25%。 為了驗證XGBR-GA模型的預測結果,本研究也依據該預測結果所對應的試片製備參數准備了NPC試片進行CO2RR。驗證試驗結果顯示由XGBR-GA模型數據所製備的試片的五大產物法拉第效率(氫氣,一氧化碳,乙烯,乙醇,乙酸) 總和為94.81%,與預測值的92.64%僅有2.34%的百分比誤差。驗證組的二碳產物法拉第效率總和則為66.71%,相對預測值的百分比誤差為2.3%。為驗證本NPC試片的工業化發展性,本研究也計算了產物時空產率,得知結果為4693 µmol hr-1g-1;與其他以純銅為基礎的試片相比高出了接近兩倍。本研究的結果也展示出了機器學習在優化CO2RR用NPC試片的巨大潛能,以便能夠在CO2RR中展現出更好的還原表現。 Since the industrial revolution, the combustion of petroleum products and other major industrial manufacturing processes has significantly contributed to carbon emissions. Apart from destabilizing the balance among ecosystems, these greenhouse gases also affect the well-being of human civilization. Since then, research projects from around the world have focused on methods to reduce the concentration of carbon dioxide in the atmosphere, with CO2 reduction reaction (CO2RR) being one of the major subjects. Among various metal catalysts used in CO2RR, nanoporous copper (NPC) acts as a highly-potential catalyst with advantages over other catalysts, such as its high reactive area and superior electrical conductivity, in addition to the ability of copper to reduce CO2 into hydrocarbons of high economic value, such as ethylene and ethanol. Adding machine learning into such research could reduce the resources needed from trial and error while producing accurate predictions. One thing to note is that current research regarding CO2RR and machine learning revolves around the optimization of catalyst design and structures, while less attention is paid to investigating the optimization of process parameters when synthesizing catalyst samples for CO2RR. Such exploration in process optimization could be an opportunity in bringing CO2RR into major industrial usages. This work is based on the CO2RR performance of NPC, where the addition of machine learning could assist and optimize NPC sample preparation to achieve the highest faradaic efficiency sum of C2 products. Data collection starts with two NPCs synthesized from copper-aluminum alloy precursors of different atomic ratios and their respective CO2RR results. The data consists of five inputs based on crucial process parameters and five outputs from major products of CO2RR. Machine learning (ML) models and genetic algorithm (GA) are combined to form an ML-GA model to determine the optimal process parameters in a designated parametric area. After comparing three ML-GA models, the combined model of extreme gradient boosting regression (XGBR) and GA (XGBR-GA model) stands out with the highest C2 product (ethylene, ethanol, and acetic acid) faradaic efficiency prediction of 68.25% without overshooting the physical limitations. A validation test is implemented to compare the actual results and the predicted values of the XGBR-GA model by applying CO2RR to the NPC sample synthesized using the predicted process parameters. The validation results show that the actual faradaic efficiency sum of five major products (hydrogen, carbon monoxide, ethylene, ethanol, and acetic acid) is recorded at 94.81%, with a percentage error of 2.34% from the predicted value of 92.64%. C2-product sum-wise, the validation test had a result of 66.71%, a percentage difference of 2.3% from the prediction. To showcase the industrial potential of the optimized NPC sample, the reaction's space-time yield (STY) was calculated to be 4693 µmol hr-1g-1, two times higher than the highest competing research using pure Cu as a base catalyst. This research proposes a promising prospect in optimizing process parameters of NPC samples for CO2RR using machine learning in the search for better reduction performances. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/88001 |
DOI: | 10.6342/NTU202301272 |
全文授權: | 同意授權(限校園內公開) |
顯示於系所單位: | 機械工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-2.pdf 目前未授權公開取用 | 8.53 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。