Localization and prediction of movement of miners in underground mines have been a constant problem more so during a mine disaster. Due to the unavailability of GPS signals, the pillars are used as a method to locate these miners, and thus, location prediction is also carried out with reference to these pillars. In this work, we demon- strate a Delay-tolerant Network (DTN) system called Miner-Finder that leverages Machine Learning (ML) framework (GAE-LSTM) that works on edge devices (e.g., mobile phones, tablets) to predict the location of miners in an underground mine. The information such as speed, angle, time, nearest pillar is first sensed by the mo- bile devices which is then sent to the GAE-LSTM framework. This framework then uses the predicted location of the miners at differ- ent times to route important messages from DTN nodes itself. For this, the system generates a routing table based on the predicted locations with their respective times and forms a contact graph for routing. The DTN system is decentralized and does not need any central/base server and the location prediction is performed locally at individual devices in a federated fashion.