Over the past decades industrial communication networks have evolved into highly diverse and heterogeneous environments, with a variety of different technologies being deployed to address the diverse requirements of manufacturing-and automation-specific use cases. These include stringent latency limits, high availability and reliability, as well as deterministic communication behavior. To assure the necessary allocation of re-sources and provisioning of required Quality-of-Service in highly diverse communication systems, a holistic network management approach is needed that can serve all cornerstones of modern industrial networks. More recently, this lead to the development of new adaptive and agile management approaches that imple-ment autonomous and self-organizing manufacturing networks, whereby Machine Learning (ML) methods started to become an integral part for overcoming the limiting factors of practically deploying such systems. Due to the growing complexity of today's networking environments, defining network management policies based on expert knowledge becomes increasingly difficult. ML has evolved as a promising technique to extract knowledge from collected data to enable cognitive network management approaches. This paper reviews past advances in ML applications for zero touch management of heterogeneous industrial communication networks. It illustrates how a network's management life-cycle that is based on digital twin technology can harnesses the potentials of ML to bring the concepts of organic computing and zero-touch cognitive manufacturing within industrial networks closer to reality. Lastly, recent papers that discuss the use of ML approaches for self-x features in Zero-Touch Management (ZTM) network environments are surveyed and relevant open issues are discussed.