Accurate prediction of Heat Release Rate (HRR) is essential for understanding fire development and improving safety strategies in buildings. HRR determines the intensity of a fire and strongly influences flame height, temperature, smoke production, and structural damage. Traditional computational fire modeling tools such as Fire Dynamics Simulator (FDS) provide detailed predictions but are computationally expensive and unsuitable for rapid or real-time analysis. At the same time, purely data-driven machine learning models often lack physical consistency and may produce unrealistic predictions when extrapolating beyond the training data. This project investigates a physics-informed machine learning approach for predicting fire Heat Release Rate in building fire scenarios. The proposed framework integrates empirical fire plume correlations with deep learning in order to combine the predictive power of data-driven models with the reliability of established fire dynamics principles. In particular, correlations developed by Gunnar Heskestad for flame height and plume behavior are incorporated into a recurrent neural network architecture based on Long Short-Term Memory (LSTM). These correlations are embedded within the learning process through physics-derived features and additional loss constraints that penalize violations of the physical relationships. Training data are generated using Fire Dynamics Simulator (FDS) through a set of parametric building fire scenarios that vary in room geometry, fuel type, ventilation conditions, and fire location. The simulations produce time-series data describing key fire variables such as Heat Release Rate, radiative heat flux, and mass loss rate. The neural network is trained to perform multi-step time-series prediction, forecasting the short-term evolution of these variables from previous fire states. The main objective of the project is to evaluate whether incorporating domain knowledge from fire science can improve predictive accuracy and physical consistency compared with purely data-driven models. Preliminary results indicate that the physics-informed model achieves improved prediction accuracy and reduces physically inconsistent outputs relative to a baseline neural network trained only on simulation data. By combining computational fire simulations with physics-guided deep learning, this work contributes to the emerging field of scientific machine learning for fire safety engineering. The proposed framework aims to provide a foundation for faster surrogate models of fire behavior that could support future applications in real-time fire monitoring, digital twins for buildings, and enhanced fire risk assessment tools.