FixFit: using parameter-compression to solve the inverse problem in overdetermined models

All fields of science depend on mathematical models. One of the fundamental problems with using complex nonlinear models is that data-driven parameter estimation often fails because interactions between model parameters lead to multiple parameter sets fitting the data equally well. Here, we develop a new method to address this problem, FixFit, which compresses a given mathematical model's parameters into a latent representation unique to model outputs. We acquire this representation by training a neural network with a bottleneck layer on data pairs of model parameters and model outputs. The bottleneck layer nodes correspond to the unique latent parameters, and their dimensionality indicates the information content of the model. The trained neural network can be split at the bottleneck layer into an encoder to characterize the redundancies and a decoder to uniquely infer latent parameters from measurements. We demonstrate FixFit in two use cases drawn from classical physics and neuroscience.