Augmented Simulation Models for the Initial Multi-physics Design of Electrical Machines

电能是现代社会运转的基础。自大约200年前电磁物理相关的基本理论发展及其在能量转换和传输中的应用以来，用于产生、传输和转换电能的系统呈指数级增长，2018年产生和消耗了约26.7太瓦时。对电能的需求增长率大约是其他能源的两倍。然而，电能并不是主要的能源——它必须由其他能源转化而来，例如化石燃料（煤、石油、天然气）、水力、风能、太阳能等。此外，电能本身不能直接使用；它需要转换成最终形式，通常是机械的或热的，以便在系统或设备中使用，如洗碗机，电动汽车或工业驱动器。电能与电能之间的转换，在很大程度上涉及到一个电磁系统，通常是发电机或电动机。每个转换和传输过程都涉及损耗，如果电机或发电机的效率为90-95%，那么从产生到使用之间的总能量损耗超过2.5太瓦时——这既代表了财务成本，也代表了对全球变暖的影响。为了减少这些损失，已经制定了法规，要求对许多驱动器实现最低效率。在许多情况下，这是通过使用电力电子驱动器以可变速度运行电机来满足的。这种变化对设计过程产生了影响，现在必须实现性能“信封”，而不是单个操作点。这种操作的变化影响了电机的设计过程，并可能使其更加复杂和昂贵。在这个过程中，越早确定诸如效率、功率因数等数量的性能信封（或地图），在探索不符合规格的设计上浪费的时间就越少。为了实现这一目标，需要一个有效的设计过程来考虑机器在整个操作范围内的完整多物理场性能。这是可以用当前的模拟工具实现的，但成本过高，通常需要数周的计算时间，本身就是巨大的能源成本。本研究的目的是开发基于机器学习代理模型的快速电机仿真器，即开发电机的黑盒模型。这将允许在移动到完整的仿真系统之前快速探索潜在的设计空间，以确定可能的设计候选，从而降低设计成本。当与增材制造系统相连接时，这种方法将能够构建更高效的机器，同时降低设备的设计和制造成本。

Electrical energy is fundamental to the functioning of modern society. Since the development of the basic theory related to electromagnetic physics and its application to energy conversion and transmission about 200 years ago, the growth in systems for generating, transmitting and converting electrical energy has been exponential resulting in about 26.7 Terawatthours being generated and consumed in 2018. The growth rate of the demand for electrical energy is about double that of other energy sources. However, electrical energy is not a primary source - it must be created by conversion from another source e.g. fossil fuels (coal, oil, gas), hydro, wind, solar, etc. Additionally, electrical energy, in itself, is not directly usable; it requires conversion to a final form, usually mechanical or thermal, in order to be used in a system or a device such as a dishwasher, an electric vehicle or an industrial drive. The conversions to and from electrical energy, to a major extent, involves an electromagnetic system often in the form of a generator or motor. Each of the conversion and transmission processes involve losses and, with a 90-95% efficiency for a motor or generator, the total energy loss between generation and usage is over 2.5 Terawatthours - and this represents both financial costs and an impact on global warming. In a step to reduce these losses, regulations have been created which require minimum efficiencies to be achieved for many drives. In many cases, this is met by operating motors at variable speed using a power electronic drive. This change has an impact on the design process where a performance "envelope" must now be achieved rather than a single operating point. This change in operation impacts the design process for an electrical machine and makes it, potentially, more complex and expensive. The earlier in the process that performance envelopes (or maps) for quantities such as efficiency, power factor, etc., can be determined, the less time is wasted in exploring designs which will not meet the specifications. To achieve this goal, an effective design process which accounts for the complete multi-physics performance of a machine over its entire operational envelope is needed. This is achievable with current simulation tools, but the costs are excessive and can often need weeks of computational time, in itself a huge energy cost. The objective of this research is to develop fast electrical machine emulators built on machine learning based surrogate models, i.e. develop black-box models of the electrical machine. These will allow a fast exploration of the potential design space to locate possible design candidates before moving to a full simulation system, thus reducing design costs. When linked to an additive manufacturing system, this approach will enable the construction of more efficient machines while reducing the design and manufacturing costs of the devices.