Crystallisation optimisation and control using surrogate modelling and adaptive model predictive control

结晶过程本身就是一个很难控制的过程，因为有许多相互作用的参数，包括成核、团聚、破碎、流体动力学等。由于潜在机制的分子性质，稳健的建模、控制和优化并不是一帆风顺的。这可能会导致批次不一致、更高的下游加工成本以及难以长期控制质量。种群平衡模型(PBM)和模型预测控制(MPC)对于这些系统是存在的，但有其局限性。PBM的计算可能很复杂，因此很难进行实时控制。由于缺乏对结晶过程中所有相互作用范围的详细了解，PBM将比实际系统的复杂性更低。该项目建议使用代理模型，如物理信息神经网络和强化学习，利用其非线性和自适应预测能力，捕获结晶系统数据中包含的所有行为范围。通过结合PBMS和实验数据对这些代理模型进行训练，可以开发出一种适用于一系列结晶系统的稳健控制和优化方法。这些模型可以实时控制和优化所需的晶粒度分布，同时通过多目标优化来平衡其他约束和目标。最终目标是一种控制系统，实现已建立的数字孪生兄弟，只需有限的实验数据。

Crystallisation is inherently a challenging process to control as a result of a number of interacting parameters including nucleation, agglomeration, breakage, hydrodynamics, and so on. Due to molecular nature of the underlying mechanisms, robust modelling, control and optimisation is not straightforward. This can result in inconsistent batches, higher downstream processing costs, and difficulties controlling quality long-term. Population Balance Models (PBMs) and Model Predictive Control (MPC) exist for these systems but have limitations. PBMs can be computationally complex and as such real-time control is difficult. Due to the lack of detailed understanding of the full scope of interactions in crystallisation, PBMs will be of reduced complexity than the real system. This project proposes the use of surrogate models such as physics informed neural networks and reinforcement learning, utilising their nonlinear and adaptive predictive capabilities, to capture the full scope of behaviour contained within crystallisation system data. By combining PBMs and experimental data in the training of these surrogate models, it is proposed that a robust control and optimisation method can be developed that can be applied to a range of crystallisation systems. The models can enable real-time control and optimisation of a desired crystal size distribution, whilst balancing other constraints and objectives through multi-objective optimisation. The final goal is a control system that realises an established digital twin, with limited experimental data required.