In this extended abstract, we present our research directions on model construction for complex systems, particularly building Systems Dynamics (SD) models from data. In our previous work, we introduced Systems Dynamics Aware Neural Networks (SDANN) that facilitate incorporating known dynamics of systems into neural network training by following a very flexible programming pattern. We also extended our work to address parameter estimation problem by considering the systems where the dynamics among variables are known in general and could be modeled with SD or differential equations but the parameter values are not known. We argued that SDANN could offer a powerful and flexible way to tackle parameter estimation problems for complex systems. Our next challenge is the application of SDANNs for model construction, that is to identify the stock variables, flows and parameters of the underlying systems dynamics model from data. Approaches exist for identifying the underlying model behavior for dynamical and physical systems, and for building the governing differential equations using available data. Main difference between existing work and our research direction is the larger complexity in identifying not just the underlying model behavior or the governing differential equation but additionally building candidate SD models with stock variables, flows and parameters that fit the data. Our research direction considers a set of pre-programmed behavior pattern classes between each pair of stock variables, and tries to find the best matching pattern from the data. Then, a set of candidate SD models are generated based on the identified behavioral patterns among the stock variables. With the application of SDANN on model construction, partially known dynamics can be incorporated via coding the difference equations or by specifying the known behavior patterns among stock variables from the set of pre-programmed behavior patterns.