另一方面，盡管大多數定性（或分類）變量（如化學元素、化學成分）比定量變量更容易獲得，但在自動材料設計中直接將定性變量作為設計變量的一部分是一個挑戰。

金屬有機框架（MOFs）就是這類材料系統的一個例子。MOFs是一類多孔結晶材料，廣泛用于氣體儲存、氣體分離和催化。由于其高度可調性，MOFs被視為解決不同應用問題的潛在方案，例如二氧化碳（CO₂）的捕集和分離。然而，由于MOF構建塊及其組合方式的多樣性，候選材料數量級過高。

因此，實驗所需的時間和資源太高，人們已經開始使用機器學習來加速材料系統的設計和開發。但現有的方法通常依賴于大量的數據集和高維物理描述符來表示材料設計空間。這些機器學習模型既耗時，泛化性又不強，通常不能遷移到不同的設計目標上。

Fig. 4 The LVGP-BO results for the Reduced Design Space (RDS) exploration.

來自美國西北大學機械系的Yigitcan Comlek等，提出了一套潛在變量高斯過程多目標批量貝葉斯優化（LVGP-MOBBO）框架，以直接從構建材料的構建塊中快速設計優越的MOFs。

他們使用了已有的定性MOFs建筑塊信息，構建了一個可解釋的LVGP模型，在MOBBO的輔助下，自適應地引導CO₂捕獲和分離性能較好的MOFs。

他們通過整合批量貝葉斯優化，無描述符的LVGP也可以有效地擴展到具有大量級別的應用。通過LVGP預測具有看不見構建塊的MOFs的特性是一個很有前途的研究領域。

該框架的一個有趣的應用是將涉及到通過自主實驗研究進行材料設計和開發。由于在LVGP-MOBBO中沒有人為干預，而且實驗輸入可以是定性和定量的，在這里提出的方法可以幫助研究人員有效地指導實驗。

Editorial Summary

With recent advances in machine learning (ML), material system design and development has undergone rapid acceleration. However, one of the major challenges in applying ML to material system design lies in finding the appropriate design representations. Most material design applications take advantage of quantitative (or numerical) design variables to represent material systems. In most cases, these quantitative descriptors (features) require either expert knowledge or data analysis to find the most appropriate ones. On the other hand, although most qualitative (or categorical) variables (e.g., chemical elements, chemical compositions) are more accessible than quantitative variables, it is challenging to directly include qualitative variables as a part of the design variables in automated materials design. Metal-organic frameworks (MOFs) are an example of such materials systems.

MOFs are a class of porous crystalline materials that have been used extensively for gas storage, gas separation, and catalysis. Because of their highly tunable nature, MOFs have been looked at as a potential solution for different applications such as CO₂ capture and separation. However, the versatility and different possible combinations of the MOF building blocks lead to millions of candidates. Due to the high experimental cost, both in time and resources, machine learning has been used to accelerate material system design and development. However, the existing approaches usually rely on large data sets and high-dimensional physical descriptors to represent the material design space. These processes can be both time consuming and property specific, meaning that the ML models and descriptors are often not transferable to different design objectives.?

Yigitcan Comlek et al. from the Department of Mechanical Engineering, Northwestern University, presented a Latent Variable Gaussian Process Multi-Objective Batch Bayesian Optimization (LVGP-MOBBO) framework to perform rapid design of superior MOFs directly from the building blocks that construct the material. They took advantage of the readily available qualitative building block information that is used to construct the MOFs and built an interpretable LVGP surrogate model that cooperates with MOBBO to adaptively lead towards promising MOF candidates for CO₂ capture and separation. With the integration of batch BO, descriptor-free LVGP can be effectively extended to applications with substantial number of levels. To predict the properties of MOFs with unseen building blocks through LVGP is a promising area of research. The interesting application of this framework would involve performing materials design and development through autonomous experimentation studies. As there is no human intervention in LVGP-MOBBO, and the experimental inputs can be both qualitative and quantitative, the method presented in this work can help researchers guide their experiments efficiently.

原文Abstract及其翻譯

Rapid design of top-performing metal-organic frameworks with qualitative representations of building blocks （快速設計具有定性表示構建塊的性能最佳的金屬有機框架）

Yigitcan Comlek, Thang Duc Pham, Randall Q. Snurr & Wei Chen

Abstract

Data-driven materials design often encounters challenges where systems possess qualitative (categorical) information. Specifically, representing Metal-organic frameworks (MOFs) through different building blocks poses a challenge for designers to incorporate qualitative information into design optimization, and leads to a combinatorial challenge, with large number of MOFs that could be explored. In this work, we integrated Latent Variable Gaussian Process (LVGP) and Multi-Objective Batch-Bayesian Optimization (MOBBO) to identify top-performing MOFs adaptively, autonomously, and efficiently. We showcased that our method (i) requires no specific physical descriptors and only uses building blocks that construct the MOFs for global optimization through qualitative representations, (ii) is application and property independent, and (iii) provides an interpretable model of building blocks with physical justification. By searching only ~1% of the design space, LVGP-MOBBO identified all MOFs on the Pareto front and 97% of the 50 top-performing designs for the CO₂?working capacity and CO₂/N₂?selectivity properties.

摘要?

定性(分類)信息的系統通常會給數據驅動材料設計帶來挑戰。特別地，通過不同的構建塊來表示金屬有機框架（MOFs）給設計者將定性信息納入設計優化帶來了挑戰，同時也帶來了一個組合型的挑戰，即設計者們能夠探索的MOFs太多。在本工作中，我們集成了隱變量高斯過程（LVGP）和多目標批量-貝葉斯優化（MOBBO），以自適應、自主和高效地識別性能最好的MOFs。我們展示了我們的方法（i）不需要特定的物理描述符，只使用構建塊來構建MOFs，通過定性表示進行全局優化，（ii）應用和屬性獨立，（iii）提供了一個具有物理證明的可解釋構建塊模型。通過僅搜索約1%的設計空間，LVGP-MOBBO識別了Pareto前沿的所有MOFs，在目前50種CO₂吸收效率與CO₂/N₂選擇性能最好的設計中搜索出了97%的樣本。