I would recommend using SciMLSensitivity with ModelingToolkit models. While there are a few things in DiffEqParamEstim that should make it compatible, specific the SciMLStructures.jl support, it doesn’t have great coverage and so manual loss functions will be a better bet for now.