Additive manufacturing (AM), which builds a single part directly from a 3D CAD model in a layer-by-layer manner, can fabricate complex component with intricate geometry in a time- and cost-saving manner.AM is thus gaining ever-increasing popularity across many industries. However, accompanied with its unique building manner and benefits thereof are the significantly complicated physics behind the AM process. This fact poses great challenges in modeling and understanding the underlying process-structure-property (P-S-P) relationship, which however is vital to efficient AM process optimization and quality control. With the advancement of machine learning (ML) models and increasing availability of AM-related digital data, ML-based data-driven modeling has recently emerged as a promising approach towards exhaustively exploring and fully understanding AM P-S-P relationship. Nonetheless, many of existing ML-based AM modeling severely under-utilize the powerful ML models by using them as simple regression tools, and largely neglect their distinct advantage in explicitly handling complex-data (e.g., image and sequence) involved data-driven modeling problems and other versatilities. To further explore and unlock the tremendous potential of ML, this research aims to attack two significant research problems: (1) from the data or pre-data-driven-modeling aspect: can we use ML to improve AM data via ML-assisted data collection, processing and acquirement? (2) from the data driven modeling aspect: can we use ML to build more capable data-driven models, which can act as a full (or maximum) substitute of physics-based model for high-level AM modeling or even realistic AM simulation? To adequately address the above questions, the current research presents a ML-based data-driven AM modeling framework. It attempts to provide a comprehensive ML-based solution to data-driven modeling and simulation of various physical events throughout the AM lifecycle, from process to structure and property. A variety of ML models, including Gaussian process (GP), multilayer perceptron (MLP), convolutional neural network (CNN), recurrent neural network (RNN) and their variants, are leveraged to handle representative data-driven modeling problems with different quantities of interest (QoI). They include data-driven process modeling (melt pool, temperature field), structure modeling (porosity structure) and property modeling (stress field, stress-strain curve). The results show that this research can break existing limitations of those five data-driven AM modeling in terms of modeling fidelity, accuracy and/or efficiency. It thus well addresses the two research questions that are key in significantly advancing data-driven AM modeling. In addition, although the current research uses five representative physical events in AM as examples, the data-driven methodologies developed should shed light on data-driven modeling of many other physical events in AM and beyond.