A Novel Framework for Automated Simultaneous Model Identification and Parameter Estimation in Kinetic Studies

数字化正在推动制造业的深刻变革，将数字孪生应用于化学反应系统设计、控制和真实的时间优化。数字孪生模型基于强大而可靠的动力学模型，可准确预测化学反应的行为，并在实验设计空间中探索各种操作条件。数字孪生模型需要进行昂贵的模型验证实验，并投入大量时间和分析资源来识别模型结构并精确识别系统特定的动力学参数集。该项目旨在开发一种新的软件框架，该框架基于物理信息机器学习（ML）技术和基于模型的实验设计（MBDoE）的集成，用于快速识别动力学模型。该框架的具体目标是：1）同时识别反应速率表达式和精确估计动力学参数; 2）在影响反应系统的不确定性和干扰下的实验条件的稳健设计; 3）最小化用于模型校准的物理运行。最终目标是将这种新的软件框架集成到UCL最近开发的反应平台中，从而允许在催化流反应器中进行自主实验（最近的视频可以在www.example.com上看到v = kMCtQqbPixk）。该项目将发展沿着四个主要步骤，每一个将持续约9个月：第1阶段。开发用于机器学习辅助MBDoE的软件模块。将开发一个软件模块，并在计算机上进行测试，实施稳健的最佳实验设计技术。在该模块中，ML模型将被集成，以模拟影响系统输入和输出的潜在不确定性和干扰。稳健的MBDoE技术的模型判别和参数精度的提高将被实施，以设计实验中存在的参数失配。最近开发的实验设计技术的动力学模型选择使用人工神经网络（ANN），目前离线应用，将应用于在线和标准MBDoE技术模型识别相比。第二阶段异常值检测模块的开发。由于从一个系统产生的数据的质量是必不可少的，以确定正确的动力学模型结构和模型参数集精确，一个模块将开发数据挖掘。这将实施基于模型的数据挖掘（MBDM）和数据驱动的离群值检测技术。这些技术将通过在所获取的数据中施加不确定性来进行计算机测试，并进行比较，以验证其在在线应用中的有效性。第三阶段开发用于物理信息ML模型识别的模块。物理信息神经网络（PINN）已经被提出来求解或发现时间相关和非线性偏微分方程（PDE）。PINN是经过训练的神经网络，用于解决监督学习任务，同时尊重特定的物理约束。与标准的深度神经网络不同，PINN在训练神经网络时将物理信息微分方程直接添加到损失函数中。PINN显示出以下优点：1）训练PINN模型仅需要少量数据; 2）通过强制遵循物理约束来保证鲁棒性; 3）直观的结果（即，PINN的输出是一组模型函数）。PINN将被集成到一个模块中，允许识别：i）反应速率表达式（即动力学模型结构）; ii）反应器模型（理想，分散模型）; iii）同时i）和ii）。第四阶段自主反应平台中软件模块的集成。配备第1、2和3阶段开发的计算模块，将在LabView中开发硬件/软件图形用户界面（GUI），用于与硬件通信。

Digitalisation is driving a deep transformation in manufacturing sectors through the application of digital twins for chemical reaction systems design, control and real time optimisation. Digital twins are based on robust and reliable kinetic models to accurately predict the behaviour of chemical reactions and explore a wide range of operating conditions in the experimental design space. Digital twin models require costly experimentation for model validation and a significant investment of time and analytical resources to identify both model structure and precisely identify the system-specific set of kinetic parameters. The proposed project aims to develop a new software framework based on the integration of physics-informed machine learning (ML) techniques and model-based design of experiments (MBDoE) for the fast identification of kinetic models. The specific goals of this framework are: 1) simultaneous identification of reaction rate expressions and precise estimation of kinetic parameters; 2) robust design of experimental conditions under uncertainty and disturbances affecting a reaction system; 3) minimisation of physical runs for model calibration. The ultimate aim is to integrate this new software framework in reaction platforms recently developed at UCL allowing autonomous experimentation in catalytic flow reactors (a recent video can bee seen in https://www.youtube.com/watch?v=kMCtQqbPixk ). The project will develop along four main steps, each of which will last ~ 9 months:Stage 1. Development of a software module for machine-learning assisted MBDoE. A software module will be developed and tested in-silico implementing robust optimal experimental design techniques. In the module, ML models will be integrated to model potential uncertainty and disturbances affecting inputs and outputs to the system. Robust MBDoE techniques for model discrimination and improvement of parameter precision will be implemented to design experiments in the presence of parametric mismatch. Recently developed experimental design techniques for kinetic model selection using artificial neural networks (ANNs), currently applied offline, will be applied online and compared with standard MBDoE techniques for model identification. Stage 2. Development of a module for outliers detection. As the quality of data generated from a system is essential to identify both the correct kinetic model structure and the set of model parameters precisely, a module will be developed for data mining. This will implement model-based data mining (MBDM) and data-driven outlier detection techniques. These techniques will be tested in-silico by forcing uncertainty in the data acquired and compared to verify their effectiveness in online applications. Stage 3. Development of a module for physics-informed ML model identification. Physics-informed neural networks (PINNs) have been proposed to solve or discover time-dependent and nonlinear partial differential equations (PDEs). PINNs are neural networks trained to solve supervised learning tasks while respecting specific physics-informed constraints. Unlike standard deep neural networks, PINNs add physics-informed differential equations directly into the loss function when training a neural network. PINNs show the following advantages: 1) training a PINN model only requires a small amount of data; 2) the robustness is guaranteed by being forced to follow physical constraints; 3) intuitive results (i.e., the output of a PINN is a set of model functions). PINNs will be integrated in a module allowing to identify: i) reaction rate expressions (i.e. kinetic model structure); ii) reactor model (ideal, dispersion models); iii) both i) and ii) simultaneously. Stage 4. Integration of software modules in autonomous reaction platforms. Armed with the computational modules developed in stage 1, 2 and 3, an hardware/software graphical user interface (GUI) will be developed in LabView for communication with the hardware.