Method and apparatus for training a system model with gain constraints. A method is disclosed for training a steady-state model having an input and an output and a mapping layer for mapping the input to the output, the model comprising a stored representation of a plant or process, and including a linear portion and a non-linear portion, where the non-linear portion includes a function. Input is received to the model, and predicted output computed corresponding to attribute(s) of the plant or process. The predicted output is stored, and is usable to manage the plant or process. The model is trained to optimize a specified objective function subject to one or more constraints, e.g., via a non-linear programming (NLP) optimizer, the constraints including, hard constraint(s) comprising strict limitations on the training in optimizing the objective function, and/or soft constraint(s) comprising a weighted penalty function included in the objective function.