Model-based control of spray synthesis of structured granules with specified properties, using transfer functions derived by multivariate stochastic models and machine learning

建议的重点是喷雾干燥的工艺链。喷雾干燥过程使载体液体雾化，载体液体包含溶解或分散颗粒形式的固体。液体在干燥器中停留时间内蒸发，通过晶体生长(溶液)或浓度增加(悬浮液)发生凝固。如此生成的区组的属性函数确定应用程序属性。在这里，孔隙率和孔结构起着核心作用，例如，对于吸附或催化行为和处理。因此，产品设计的工作就是裁剪颗粒的内部形态。后者取决于进料的性质，例如颗粒、它们的相互作用、固体浓度，以及工艺条件，例如温度场、液滴大小、干燥条件。为了控制整个过程，必须考虑所有这些参数，从而导致多参数问题。从在线测量中获得的工艺数据对于所需的最终产品规格是有限的。有必要使用计算机断层扫描(CT)等离线高分辨率方法来获得关于颗粒孔结构的完整信息。这是耗时的，从采样、样品准备到图像分割和数据分析都要经历几个步骤。为了能够在控制回路中使用CT数据，它必须连接到串联数据，这是使用代理来完成的。代理是从图像数据推导出的单个或组合参数。将使用两种代理策略：可解释的代理(例如，孔隙属性分布)和数据驱动的代理(例如，通过自动编码获得的颗粒系统的压缩图像数据)。通过机器学习的方法，如前馈网络，实现了过程参数和饲料特性与可解释的和数据驱动的代理之间的正向和反向连接。在第一个资助期(FP)中，讨论了液滴干燥过程中的喷雾生成和颗粒结构的连接步骤。这将在第二个FP中通过定义的预处理步骤来扩展，例如，在进料流中定制颗粒-颗粒相互作用和解聚(机械剪切/超声波)。然后，控制概念获得更多的自由度，因为馈送的预结构化成为可能。在第一FP中开发的基于模型的控制方案通过在从粒子系统的分布属性或应用自动编码的概念抽象地集成的图像数据(在线动态成像/离线CT图像)获得的代理之间执行统计学习来使用传递函数。考虑到来自其他工艺步骤的类似数据，它必须在第二个FP中进行扩展。通过这种方式，该项目利用在线和离线信息，为喷雾干燥范例产生了对数据驱动的黑匣子控制方案的性能和适用性的科学见解。

The proposal focusses on the process chain of spray drying. A spray drying process atomizes a carrier liquid, which contains solids in the form of dissolved or dispersed particles. The liquid evaporates during the residence time in the dryer and solidification occurs, either by crystal growth (solutions) or by concentration increase (suspensions). The property function of the so-generated granules determines the application properties. Here, porosity and pore structure play a central role, e.g. for adsorptive or catalytic behavior and handling. Therefore, product design works on tailoring the inner morphology of granules. The latter depends on both, the properties of the feed, e.g. the particles, their interactions, the solid concentration, and the process conditions, e.g. temperature field, drop size, drying conditions. To control the entire process all these parameters have to be considered leading to a multiparametric problem. The process data available from in-line measurements is limited concerning the desired final product specification. It is necessary to employ off-line high-resolution methods like computer tomography (CT) to acquire the full information on the pore structure of the granules. This is time consuming and undergoes several steps from sampling, sample preparation to image segmentation and data analysis. To be able to use CT-data in the control loop it has to be connected to in-line data, which is done using proxies. A proxy is a single or combined parameter deduced from image data. Two strategies for proxies will be used: interpretable proxies (e.g. pore property distributions) and data-driven proxies (e.g. compressed image data of particle systems obtained by automated encoding). The forward and backward connection of process parameters and feed properties with interpretable and data-driven proxies is achieved by methods of machine learning, e.g., feed-forward networks.In the 1st funding period (FP) the connected steps of spray generation and particle structuring during droplet drying are addressed. This will be extended in the 2nd FP by defined preconditioning steps, e.g., tailoring of particle-particle-interactions and de-agglomeration (mechanical shear / ultrasound) within the feed stream. Then the control concept obtains more degrees of freedom, since a pre-structuring of the feed becomes possible. The model-based control scheme developed in the 1st FP uses transfer functions by performing statistical learning between proxies derived either from distributed properties of particle systems or abstractly integrated image data (in-line dynamic imaging / off-line CT-images) applying the concept of automated encoding. It has to be extended in the 2nd FP considering similar data from additional process steps. In this way, the project generates for the example spray drying scientific insights on the performance and applicability of a data-driven black-box control scheme, using both in-line and off-line information.