DEVELOPMENT OF A MACHINE LEARNING-ASSISTED DIGITAL TWIN PLATFORM FOR REAL-TIME OPTIMISATION OF REACTION SYSTEMS UNDER UNCERTAINTY

波音777双引擎喷气式飞机于1995年投入使用，是世界上第一架100%数字化设计的飞机。计算机辅助设计被证明比人类工程团队更精确，所有未来计划的实体模型都被取消了。尽管这种情况几十年前就发生在航空业，但在化学工业中还没有被复制，尽管化学和制药行业是英国经济中第三大制造业。这是这个项目渴望实现的愿景：在不需要物理原型的情况下，以数字方式设计一个化工厂。根据工业4.0范式，该项目旨在开发一个“数字孪生”平台，在该平台上，基于可重构数学模型的化学过程的硅替代品被用于快速探索替代和创新的解决方案，以设计新的可持续过程，以及对化学过程的强大模拟、控制和优化，以实现净零排放等可持续发展目标。然而，只有在基础模型提供了用于放大模型的反应系统的准确描述的情况下，才能获得适用于广泛操作条件探索的可靠数字孪生。确定合适的数字孪生模型需要在实验和分析资源以及人力方面进行大量投资，以开发和严格验证预测模型。为了使反应建模研究更便宜，更快，更适用于工业，我们打算通过开发数字孪生平台技术，在制药和精细化学品制造方面带来相当大的变化，其中自动化，人工智能和实验算法优化设计的好处被合并，用于快速识别预测多保真模型，包括基于物理的模型和代理机器学习（ML）模型。该平台将结合数字孪生软件，其中虚拟测试，先进的物理信息ML和最佳实验设计算法用于快速决策，灵活的反应器（Taylor-vortex反应器）可以保证有效的质量和热传递以及可调节的流体动力学。使用泰勒涡反应器的动机是，由于缺乏设计指南和跟踪记录，它是一种反应器类型，在化学工业中采用有限，尽管它为制造业提供了一个现实的选择。因此，它提供了一个很好的例子来证明数字孪生技术在降低化学过程开发和扩大风险方面的力量。通过这些算法识别的计算成本低廉的代理ML模型将推动实验的在线设计和实时优化，允许在没有用户干预的情况下操作平台，并能够快速生成信息数据集，快速识别动力学，质量和传热模型，同时对时间，人力和分析资源的影响最小。为了开发该平台并确保其在工业领域的直接适用性，我们与大型制药公司葛兰素史克（GSK）和两家大型化工公司巴斯夫（BASF）和强生（Johnson & Matthey）建立了直接合作伙伴关系，以确保知识的转移和开发的平台对化学和制药制造的直接影响。该团队由Autichem（设备供应商）和Quotient Sciences（药物开发和制造加速器）补充，这两家中小企业在整个药物开发途径中与全球制药公司合作，以协助开发新的制造工艺和方法。公司将参与指导研究，并确保其成果与工业相关，并最终在工业研发和化学和制药工艺制造中得到利用。

The Boeing 777 twin engine jet that entered service in 1995 was the world's first 100% digitally designed aircraft. The computer-aided design was proven to be more accurate than a human engineering team could be and all future planned physical mock-ups were cancelled. Even though this happened decades ago in the airline industry, this has not yet been replicated in the chemical industry, despite the chemicals & pharmaceuticals sector being the 3rd largest manufacturing sector in the UK economy. This is the vision that this project aspires to contribute to: design a chemical plant digitally without the need for physical prototypes. In line with the Industry 4.0 paradigm, this project aims to the development of a "digital twin" platform where in-silico surrogates of chemical processes based on reconfigurable mathematical models are used to quickly explore alternative and innovative solutions for the design of new sustainable processes, and for the robust simulation, control and optimisation of chemical processes, to achieve sustainability targets such as net-zero emissions. However, reliable digital twins, suitable for the exploration of a wide range of operating conditions, can be obtained only if the underlying models provide an accurate description of the reaction systems to be used in scale-up models. The identification of suitable digital twin models requires a significant investment in terms of experimental and analytical resources, as well as manpower to develop and rigorously validate predictive models. To make reaction modelling studies cheaper, faster and more industrially applicable, we intend to bring about a sizable step change in both pharmaceuticals and fine chemicals manufacturing by developing a digital twin platform technology, where the benefits of automation, AI and optimal design of experiments algorithms are merged for the quick identification of predictive multifidelity models, including physics-based models and surrogate machine learning (ML) models. The platform will combine a digital twin software, where virtual testing, advanced physics-informed ML and optimal experimental design algorithms are used for fast decision-making, with flexible reactors (Taylor-vortex reactors) that can guarantee efficient mass and heat transfer and adjustable hydrodynamics. The use of Taylor-vortex reactors is motivated by the fact that it is a reactor type with limited adoption in the chemical industry due to the lack of design guidelines and track record, even though it provides a realistic option for manufacturing. Thus, it provides an excellent exemplar to demonstrate the power of digital twin technology in derisking chemical process development and scale-up.Computationally cheap surrogate ML models identified by these algorithms will drive the online design of experiments and real time optimization, allowing to operate the platform without user intervention and enabling the fast generation of informative data sets and the quick identification of kinetics, mass, and heat transfer models with minimum impact on time, human and analytical resources. In order to develop this platform and ensure its direct applicability in the industrial sector, we have as direct collaborators a large pharmaceutical company, GSK, and two large chemical companies, BASF and Johnson & Matthey, to ensure transfer of knowledge and direct impact of the developed platform on chemical and pharmaceutical manufacturing. The team is complemented by Autichem (equipment provider) and Quotient Sciences (drug development and manufacturing accelerator), two SMEs who work with global pharmaceutical companies across the entire medicine development pathway to assist the development of novel manufacturing processes and approaches. Companies will contribute to directing the research and ensuring its outcomes are industrially relevant and eventually exploitable in industrial R&D and in chemical and pharmaceutical process manufacturing.