Scientific workflows have become a popular computational model in a variety of application domains, such as astronomy, material science, physics, and biology. As scientific applications are moving to the cloud to take advantage of the elasticity and service level agreement of resources, there has been a number of recent research efforts on cloud-based workflow systems that support various types of performance guarantees under resource cost constraints. However, most of the related work often requires advanced knowledge about workflow structures to perform scheduling and resource optimization. In addition, existing workflow systems usually employ a monolithic approach in workflow implementation and execution, which makes them inefficient in dealing with heterogeneous types of workflows. In this paper, we present MONAD, a self-adaptive micro-service infrastructure for heterogeneous scientific workflows. Specifically, our micro-service architecture helps improve the flexibility of workflow composition and execution, and enables fine-grained scheduling at task level, considering task sharing across different workflows. In addition, we employ a feedback control approach with artificial neural network-based system identification to provide resource adaptation without any advanced knowledge of workflow structures. Our evaluation on multiple realistic heterogeneous workflows demonstrates that our system is robust and efficient in dealing with dynamic scientific workloads.