CDS&E: On-Line Extraction and Separation Method Development using Surrogate Optimization and Machine Learning

在化学系化学测量和成像计划的支持下，Kevin Schug和他的分析化学小组将与Victoria Chen、Shouyi Wang、Chen Kan和Jay Rosenberger与德克萨斯大学阿灵顿分校的随机建模、优化和统计中心(COSMOS)合作，探索用于在线分析仪器的高级优化程序。痕量化学分析通常要求事先准备样品，以选择性地提取某一类化合物或减少干扰。手工样品制备是常见的，但与自动化方法相比可能会引入误差。舒格集团专注于利用超临界流体萃取和层析等强大方法的自动化在线样品制备和化学分析平台。该分析系统非常广泛地适用于各种化合物的测量。然而，有许多变量需要优化，它们对测量结果的综合影响很难合理化。为了应对这一挑战，该团队正在开发一种基于机器学习的代理优化程序，以有效地探索最佳条件。代理优化运行的结果将使用新的多任务学习来建模，以便为近似最佳检测条件提供一般知识。参与该项目的研究生和本科生将在高级数据处理和分析测量方面获得补充经验。这项工作的结果将在地方、地区和国家会议以及同行评议的出版物中广泛传播。在线超临界流体萃取(SFE)与超临界流体色谱(SFC)相结合，在非常广阔的应用空间提供了自动化的样品制备和化学分析。该团队正在整合统计学、机器学习和运筹学，以支持SFE-SFC方法的开发。SFE-SFC，特别是与质谱学(MS)检测相结合，具有几乎通用的对固体样品中所含小分子的综合分析能力。为了实现这些目标，该团队正在(1)探索样品材料、分析物和SFC柱化学等输入特征的创新表示；(2)开发符合SFE-SFC系统约束的输入特征空间和实验参数空间的新实验设计(DoE)过程；(3)在严格的多目标代理优化中实施DOE过程，该优化结合了来自混合整数、线性和二次规划的可证明最优的技术；(4)将非参数贝叶斯建模与高斯过程相结合，推导出“小数据”多任务学习算法，为输入特征空间上的优化参数构造灵活的多响应预测模型。SFE-SFC有可能取代其他离线、效率较低、绿色程度较低的方法。这项研究旨在极大地减轻那些实施SFE-SFC的人未来的负担。此外，为“小数据”开发的能源部、机器学习和代理优化方法将普遍适用于医疗保健和“绿色”建筑等领域的广泛社会挑战，这些领域的患者或建筑样本通常有限。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

With support from the Chemical Measurement and Imaging Program in the Division of Chemistry, Kevin Schug and his analytical chemistry group will collaborate with Victoria Chen, Shouyi Wang, Chen Kan, and Jay Rosenberger with the Center on Stochastic Modeling, Optimization, & Statistics (COSMOS) at the University of Texas at Arlington to explore advanced optimization procedures for on-line analytical instrumentation. Trace chemical analysis often requires that a sample be prepared beforehand to selectively extract a certain class of compounds or to reduce interferences. Manual sample preparation is common, but can introduce error relative to an automated approach. The Schug group focuses on an automated on-line sample preparation and chemical analysis platform utilizing powerful methods known as supercritical fluid extraction and chromatography. This analytical system is very broadly applicable to the measurement of a wide variety of chemical compounds. However, there are many variables that need to be optimized, and their combined effect on the outcome of the measurement is difficult to rationalize. To address this challenge, the team is developing a machine learning-based surrogate optimization procedure to efficiently explore optimal conditions. The results of the surrogate optimization runs will be modeled using novel multi-task learning for “small data,” so as to provide general knowledge for approximating optimal instrumentation conditions. The graduate and undergraduate students involved in this project will gain complementary experience in advanced data handling and analytical measurements. Results of the work will be widely disseminated at local, regional, and national conferences, and in peer-reviewed publications.Online supercritical fluid extraction (SFE) coupled with supercritical fluid chromatography (SFC) offers automated sample preparation and chemical analysis over a very broad application space. The team is integrating statistics, machine learning, and operations research to enable SFE-SFC method development. SFE-SFC, especially combined with mass spectrometry (MS) detection, has near-universal capabilities for comprehensive analysis of small molecules contained in solid samples. In pursuit of these aims, the team is (1) exploring innovative representations for input features such as sample material, analyte, and SFC column chemistry; (2) developing new experimental design (DoE) processes over the input feature space and over the experimental parameter space that comply with the constraints of the SFE-SFC system; (3) implementing the DoE process within a rigorous multiple-objective surrogate optimization that incorporates provably-optimal techniques from mixed integer linear and quadratic programming; and (4) deriving "small data" multi-task learning algorithms integrating nonparametric Bayesian modeling with Gaussian processes to construct a flexible multi-response predictive model for the optimized parameters over the input feature space. SFE-SFC has the potential to supplant other methods that are off-line, less efficient, and less “green.” This research aims to vastly reduce the future burden for those who implement SFE-SFC. Further, the developed DoE, machine learning, and surrogate optimization methodologies for "small data" will be generally applicable and available for a wide range of societal challenges in domains such as healthcare and “green” building, for which the sample of patients or buildings is typically limited.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.