State- and Parameter-space Exploration and Process Optimisation

该项目的目标是减少优化多阶段生产工艺参数所需的样品数量。在这个项目中，除了最先进的技术之外，还利用了制造工艺的以下常见特性。(I)多阶段生产过程允许观察早期的子过程及其结果，这可用于有针对性地在线优化后期阶段的参数。为了最大限度地利用现有样本，将贝叶斯优化算法扩展到多阶段设置中，贝叶斯优化算法可以集成概率先验模型和专家知识。(2)有限维过程参数也可以描述前馈控制的轨迹。在这里，参数之间的空间关系为指导优化的改进的代理模型提供了额外的洞察力。(Iii)稳健的探索和优化考虑了过程的随机性，并限制了结果偏离定义的允许集的概率。制造工艺参数优化的一个典型应用是防止对物理工艺硬件的损坏，并限制勘探和优化导致的不可用废品的可能性。(4)与参数空间中的探索不同，状态空间探索有目的地引导我们到达制造过程或结果产品的目标状态。为此，理想的做法是使用反向代理模型，该模型将目标状态映射到匹配的过程参数。结合状态空间和参数空间优化，我们可以利用这两种方法的不同维度和敏感度(相对于它们的优化目标)。在多阶段设置中，能够有目的地实现定义的中间状态(对于不同的初始条件)有助于探索和优化后续过程阶段。

The objective of this project is the reduction of the number of samples required for the optimisation of the parameters of multi-stage production processes. Beyond the state of the art, the following common properties of manufacturing processes are exploited in this project. (i) Multi-stage production processes allow the observation of early subprocesses and their results, which can be used for targeted online optimisation of the parameters of later stages. To make the best use of the available samples, Bayesian Optimisation – which can integrate probabilistic prior models and expert knowledge – is extended for the multi-stage setting. (ii) Finite-dimensional process parameters can also describe trajectories for feedforward control. Here the spatial relationship between the parameters gives additional insight to be used for improved surrogate models that guide the optimisation. (iii) Robust exploration and optimisation takes the stochasticity of the process into account and bounds the probability of the results deviating from a defined admissible set. A typical application in manufacturing process parameter optimisation is to prevent damage to the physical process hardware and to bound the probability of unusable scrap products resulting from the exploration and optimisation. (iv) In contrast to exploration in the parameter space, state-space exploration purposefully steers us to a target state, of the manufacturing process or of the resulting product. For this, one ideally uses an inverse surrogate model that maps the target state to matching process parameters. Combining state- and parameter-space optimisation allows us to exploit the different dimensionalities and sensitivities (with respect to their optimisation targets) of both approaches. In the multi-stage setting, being able to achieve, purposefully, a defined intermediate state (for varying initial conditions) helps in the exploration and optimisation of later process stages.