The conventional approach to developing Large Eddy Simulation (LES) sub-grid-scale (SGS) combustion models lacks effectiveness due to reliance on numerous tunable parameters, which vary with different flow and combustion configurations. In response, we propose leveraging machine learning to train deep neural networks using high-fidelity direct numerical simulation (DNS) combustion databases for learning LES combustion models. By eliminating adjustable parameters, our approach aims to achieve superior performance with reduced computational costs. We will explore various neural network architectures to construct LES combustion models and investigate strategies for embedding physical constraints. The performance of the trained models will be assessed using different DNN architectures, and the well-learned DNN combustion model will be utilized to propose enhancements for existing algebraic models. To validate our approach, we will evaluate the prediction accuracy of the learned LES models using new DNS databases and in posterior LES simulations. This project represents a significant step toward improving LES combustion modeling through physics-informed machine learning methods.