SEAA 2025: Supplementary Artifact

Machine Learning (ML) is increasingly integrated into software systems, introducing a set of recurring architectural challenges not commonly addressed in traditional development, such as managing data-driven components, uncertainty, and quality concerns specific to ML. While established software design patterns exist for conventional systems, a coherent set of patterns for ML-enabled architectures is still lacking. Such patterns would help guide recurring decisions and make their trade-offs more transparent to architects.

This paper presents 14 design patterns identified from a set of 49 component models of ML-enabled systems, which we compiled through a multivocal literature review covering both academic and grey literature sources. Each pattern captures a recurring architectural decision grounded in real-world practice.
To assess their practical relevance and implications, we conducted interviews with 10 experts, focusing on how the identified patterns impact key quality attributes such as maintainability, reliability, explainability, and fairness.
The resulting pattern collection supports software architects in reasoning about trade-offs in ML-based system design and provides a foundation for further research on architectural best practices.

机器学习（Machine Learning，ML）正日益与软件系统深度融合，由此带来了一系列在传统软件开发场景中鲜有被覆盖的架构共性挑战，诸如数据驱动组件的运维管理、不确定性应对，以及机器学习专属的质量管控议题。尽管传统软件系统已具备成熟的设计模式体系，但面向机器学习赋能架构的系统化设计模式仍尚未形成统一且完备的集合。此类模式可为架构师的重复性决策提供指引，并使决策中的权衡考量更为清晰透明。

本文通过涵盖学术文献与灰色文献的多源文献综述，整理得到49个机器学习赋能系统的组件模型，并从中提炼出14种设计模式。每一种模式均对应一项扎根于真实行业实践的共性架构决策。

为评估这些模式的实际应用价值与落地影响，我们采访了10位领域专家，重点探讨所提炼的模式对可维护性、可靠性、可解释性及公平性等核心质量属性的影响机制。

最终形成的设计模式集能够辅助软件架构师对机器学习驱动系统设计中的权衡问题进行推演论证，并为后续架构最佳实践的相关研究提供了重要的研究基础。