| 摘要: | 金屬有機框架多孔材料憑藉其優越的多孔性和功能性，被認為是極具潛力的氣體吸附劑。為了更有效率地篩選出最佳的氣體吸附劑，可以將分子模擬和機器學習結合，從而加速有潛力的多孔材料的開發。傳統上在機器學習中的多孔材料描述符主要依賴於該材料的結構和化性，這導致需要進一步運算每個多孔材料特性，然而由於金屬有機框架是模組化的結構，透過選擇不同的金屬節點和有機配體來建構多孔材料，因此多個金屬有機框架是共享相同的有機分子配體，並且能重複使用該資訊在不同多孔材料上。因此，此項研究發表了新的模組化的有機配體的表面電荷分佈描述符，表面電荷分佈稱為sigma profile，是使用導電類似溶劑模型進行生成，該模型結合了溶劑熱力學和計算量子力學，它關注於分子表面片段的概率分佈，考慮其特定的電荷密度。
這項研究結合分子模擬和機器學習，來預測碳氫氣體及酸性氣體兩種類型在金屬有機框架材料中的吸附特性。首先，利用蒙地卡羅Widom插入方法來計算亨利常數（KH），該常數表示金屬有機框架中碳氫類和酸性氣體的吸附強度。再者，建立了兩種類型機器學習模型包括隨機森林（RF）和極限梯度提升（XGBoost）來預測在3180個金屬有機框架材料中甲烷、乙烷、丙烷、乙烯、二氧化碳和硫化氫等氣體的吸附行為，由兩種方法確認較精準的學習模型，並利用此模型探討sigma profile描述符幫助模型性能的提升。本研究進一步隨機選擇100種由兩種有機配體組成的MOF對碳氫氣體進行模型測試，sigma profile的加入確實得到更高的預測精準度。此外，基於特徵重要性分析證實sigma profile模組化描述子確實起了相當重要的作用。最後，探討減少sigma profile特徵數量對預測精準度的影響顯示，誤差的增加可以忽略不計，整體效能依然穩定，也就是說即便使用較少的 sigma profile 特徵，也能在有限的計算資源下有效進行預測，也再次印證模組化描述符在金屬有機框架中的可用性。
總體而言，此項研究引入一種新穎的描述符，採用基於模組化的描述符特徵來構建精準的機器學習模型。這項研究的發現有望幫助未來識別用來各種吸附應用的多孔材料吸附劑。而本論文所構建的sigma profile數據庫可提供作為多孔材料領域的學者的寶貴資源。
;Metal-organic frameworks (MOFs) are considered highly promising gas adsorbents due to their superior porosity and functionality. To more efficiently screen for the best gas adsorbents, molecular simulation and machine learning can be combined to accelerate the development of potential porous materials. Traditionally, descriptors in machine learning for porous materials mainly rely on the structure and chemical properties of the material. This requires further calculations for each porous material. However, since MOFs are modular structures built by selecting different metal nodes and organic linkers, multiple MOFs share the same organic linkers, allowing the reuse of this information across different porous materials.
This study introduces a new modular descriptor for the surface charge distribution of organic linker, known as the sigma profile. The sigma profile is generated using the conductor-like screening model, which combines solvent thermodynamics and computational quantum mechanics. It focuses on the probability distribution of molecular surface fragments, considering their specific charge densities. This study combines molecular simulation and machine learning to predict the adsorption characteristics of two types of gases, hydrocarbon and acid gases, in MOFs. First, the Henry’s coefficient (KH), which indicates the adsorption strength of hydrocarbon and acid gases in MOFs, is calculated using the Monte Carlo Widom insertion method. Next, two types of machine learning models, random forest (RF) and extreme gradient boosting (XGBoost), are established to predict the adsorption behaviors of methane, ethane, propane, ethylene, carbon dioxide, and hydrogen sulfide in 3180 MOF materials. The most accurate model is identified, and the enhancement of model performance by the sigma profile descriptor is explored. The study further tests the model on hydrocarbons in 100 MOFs composed of two types of organic linkers, demonstrating that the sigma profile significantly improves prediction accuracy. Feature importance analysis confirms the significant role of the sigma profile descriptor. Finally, the impact of reducing the number of sigma profile features on prediction accuracy is investigated, showing that the increase in error is negligible and overall performance remains stable. This indicates that even with fewer sigma profile features, effective predictions can be made with limited computational resources, reaffirming the utility of modular descriptors in MOFs.
Overall, this study presents the sigma profile, a novel modular descriptor for predicting gas adsorption in MOFs using molecular simulation and machine learning. The sigma profile enhances prediction accuracy and efficiency, offering a valuable resource for future research. |
