Hybrid Physics-Neural Network Soft Sensors for Dynamic Operation of Liquid-Liquid Separation Processes

该项目的目标是探索物理信息神经网络的潜力，用于推断液-液分离过程中难以测量的条件，其中机制描述仅部分可用。作为具体的例子，我们研究了一种用于分离液-液分散体的卧式、连续操作的重力沉降器，这是化学、生物技术和回收过程中常见的低能量分离装置，其动态操作不能仅仅基于机械模型来描述。我们的目标是通过基于物理的正则化，将机器学习与机械建模相结合，开发一种可靠的软传感器来测量分散层高度，即沉降器中的主要分离性能指标。为了生成训练数据，我们将以闭环连续模式在技术规模上操作重力沉降器，该模式允许改变和测量操作条件、材料系统、分散和相分离参数。我们将混合物理-神经网络与完全数据驱动的基准进行比较，例如，循环神经网络，以验证我们的期望，即混合模型需要更少的训练数据，更好的泛化，并做出更物理一致的预测。最后将对混合软传感器进行控制演示，并对模型的有效性范围进行评估。我们的项目解决了SPP 2331中定义的三个核心挑战，即在ML模型中引入物理定律，最佳决策制定以及增加对ML应用程序的信任。

The goal of this project is to explore the potential of physics-informed neural networks for inferring poorly measurable conditions from liquid-liquid separation processes for which mechanistic descriptions are only partially available. As specific example, we investigate a horizontal, continuously operated gravity settler for the separation of liquid-liquid dispersions, a common low-energy separation unit in chemical, biotechnological, and recycling processes, whose dynamic operation cannot be described solely on the basis of mechanistic models. We aim to develop a reliable soft sensor for the dispersion layer height, i.e., the principal separation performance indicator in the settler, by combining machine learning with mechanistic modeling through physics-based regularization. To generate training data, we will operate a gravity settler on a technical scale in a closed-loop continuous mode that allows to vary and measure operating conditions, material system, dispersion, and phase separation parameters. We will compare the hybrid physics-neural network to a fully data-driven benchmark, e.g., a recurrent neural network, to validate our expectation that the hybrid model requires less training data, generalizes better, and makes more physically-consistent predictions. A demonstration of the hybrid soft sensor in control and an assessment of model validity range will conclude the work. Our project addresses three central challenges defined in the SPP 2331, namely, the introduction of physical laws in ML models, optimal decision making, and increasing trust in ML applications.