I implemented an explicit, age-structured population model for sockeye salmon runs that could be modified to include dynamics of varying complexity. Survival, smoltification, and spawning probabilites at each age and time step were modeled as constants and as time-varying functions of environmental covariates using modular neural networks. The goal of this approach was to allow the neural network components of the model to fit the unknown biological rates, while retaining the mechanistic structure of an age-structured population dynamics model. Ideally, this should allow for model-fitting flexibility while reduce the number of parameters required and the chances of overfitting. I also used non-mechanistic recurrent neural networks as an alternative model to provide a point of comparison.