Subproject SP7: Self-learning control of the catalytic conversion of olefins to α-amino acids and β-amino alcohols

该提案是FOR5538研究单元的一部分。在这个子项目中，我们的目标是开发用于在线优化烯烃催化转化为α-氨基酸和β-氨基醇的自学习控制系统。这个子项目补充了SP6中进行的综合分子、材料和工艺设计。它解释了工厂模型不匹配和不可预见的干扰，通过在单批和/或批对批水平上重复在线优化，使用可用的测量信息和所谓的混合数学模型自动重新调整操作变量。混合建模将其他子项目中获得的物理化学知识与机器学习的数据驱动方法相结合。这一办法的基本步骤包括：(i)单个过程步骤的混合数学模型，（ii）适合在线模型自适应的方法，（iii）在线优化的有效策略，（iv）将上述方法集成到自学习控制概念中(v)对所开发方法进行系统的计算机测试，最后（vi）与本研究单元所涉及的相应子项目合作进行实验验证。除了非线性模型外，还将考虑局部线性化过程描述族，并将其与非线性方法进行比较。为了在存在不确定性的迭代优化过程中安全地满足关键操作约束，提出了一种直接控制下层关键变量的分层方法。第一期重点研究α-酮羧酸酶法转化α-氨基酸，特别是同苯丙氨酸，α-羟基酮酶法转化β-氨基醇，特别是同苯丙醇，产物结晶一体化，SP3研究，随后研究催化剂回收和溶剂分离的膜分离，SP4研究。由于这个子项目的方法也依赖于SP3中产生的先验知识，在启动阶段，酶反应结晶过程的方法将首先开发并测试一个模型系统，该模型系统已在本研究单元的准备工作中研究过。第二个供资期的可能延长包括将进一步的反应步骤纳入全面的自我学习控制战略，特别强调“全厂”方面，即必须以何种方式控制哪些步骤以实现整个过程目标。此外，我们的目标是将SP6中的设计方法与本子项目中考虑的控制方法集成在一起。

This proposal is part of the research unit FOR5538. In this subproject we aim at the development of self-learning control systems for online optimization of catalytic conversions of olefins to α-amino acids and β-amino alcohols. This subproject complements the integrated molecular, material and process design performed in SP6. It accounts for plant-model mismatch and unforeseen disturbances by repetitive online optimization on the single batch and/or the batch-to-batch level to automatically re-adjust the operational variables using available measurement information and so-called hybrid mathematical models. Hybrid modeling combines physical chemical knowledge obtained in the other subprojects with data-driven approaches from machine learning. Essential steps of this approach comprise the development of: (i) hybrid mathematical models of the individual process steps, (ii) suitable methods for online model adaption, (iii) efficient strategies for online optimization, (iv) the integration of the aforementioned methods into self-learning control concepts (v) systematic in silico testing of the developed methods, and finally (vi) their experimental validation in cooperation with the corresponding subprojects involved in this research unit. Besides nonlinear models also families of locally linearized process descriptions will be considered and compared to the nonlinear approach. To safely satisfy crucial operational constraints during the iterative optimization in the presence of uncertainties, a hierarchical approach with direct control of critical variables on the lower level is suggested. Focus in the first funding period will be on the enzymatic conversion of α-keto carboxylic acids to α-amino acids, especially homophenylalanine, and α-hydroxy ketones to β-amino alcohols, especially homophenylalaninol with integrated product crystallization as studied in SP3, and afterwards on membrane separation for catalyst recycle and solvent separation, as studied in SP4. Since the approach of this subproject relies also on a priori knowledge to be generated in SP3, during the startup phase, the methodology for the enzymatic reactive crystallization process will be first developed and tested for a model system which has been studied in the preparatory work of this research unit. Possible extensions for a second funding period comprise the integration of further reaction steps into an overall self-learning control strategy with special emphasis on ‘plantwide’ aspects, i.e. which steps have to be controlled in which way to achieve the overall process objectives. Further, we aim at an integration of the design approach in SP6 with the control approach considered in this subproject.