After fire is discovered in an underground coal mine, a decision must be made to mitigate fire consequences. The decision should be made based on existing conditions, with the goal of increasing the probability of fire extinguishing without compromising the health and safety of the firefighting personnel. However, the determination of fire conditions can be difficult due to coarse in-situ measurements, fire hazards, and the large domains of interest. Additionally, CFD and network models used for predicting fire conditions are computationally expensive with long simulation processing times for informing real-time decision making. A new generalized procedure to design artificial neural networks (ANNs) capable of making predictions of fire conditions, performing hazard/risk assessment, and providing useful information to the firefighters is presented and applied to different underground coal mine fire scenarios. The feed-forward ANNs were developed to classify fires so as to provide the best firefighting decision and determine useful information in real time, such as response time and fire size. The networks were trained to make predictions on different mine locations and to use only available and measurable information in underground coal mines as inputs. The data used for training and testing the networks was generated using high-fidelity CFD and network fire simulations. Additionally, this research presents the applicability of optical fiber sensing technology for continuous, distributed, and real-time sensing. This new technology could be used for collection of input parameters during ongoing fires, leading to improvement of the prediction performance of the ANNs developed. Finally, a new approach to simulate firefighting foam flow through gob areas is proposed and tested using experimental results obtained from a scaled down experimental setup. / Doctor of Philosophy / Mine fires still represent a serious hazard in underground coal mines. The MSHA incident database shows that around 600 mine fire incidents and 33 fatalities were reported in the U.S. during the last two decades. Most fatalities and injuries that occurred in the aforementioned incidents can be attributed to lack of knowledge on existing fire conditions, leading to poor subjective decisions during fire response. Unfortunately, the in-situ determination or prediction of fire conditions are not easy tasks due to fire hazards, mine entries extensions, and simulation processing times. For this reason, this work presents new data-driven models capable of predicting and evaluating fire conditions. Its goal is to recommend the most suitable firefighting decision, as well as determine fire characteristics and response time to increase the probability of fire extinguishing without compromising mine personnel health and safety. These data-driven models are composed of artificial neural networks (ANNs), allowing for performing predictions in real time and using only available information in underground coal mines. The data used for training and testing these ANNs was generated from fire simulations. Additionally, this research proposes a new technology, such as optical fiber sensing for continuous, distributed, and real time sensing. Optical fiber sensing could contribute with more precise ANNs inputs collection, leading to a better performance prediction. Finally, an alternative way to simulate firefighting foam through gob areas for fire mitigation was proposed and tested using results obtained from experiments. This work represents a significant advancement in underground coal mine fire characterization and response.