Learning nonlinear, history-dependent material constitutive laws is a challenging task. In the local state approach with internal variables, additional unobservable variables are introduced to describe the material state. These variables are typically chosen in advance based on prior material knowledge, homogenisation or implicitly extracted from data using machine learning tools. However, this choice is not unique and may result in overly complex/simplistic representations i.e. model bias. This work explores the concept of independence between internal variables to discover an experimentally interpretable parametrization of the mechanical state. Here, independence refers to the idea that internal variables are generally associated with dissipative mechanisms that evolve independently. In machine learning, the search for independent representations is related to the so-called Blind Source Separation (BSS) problem, which, in the linear case, is often addressed using Independent Component Analysis (ICA) [1]. This method is applied to a thermoelastic material model to recover certain original sources that were intentionally “forgotten”, using a neural network autoencoder. The autoencoder replicates the behaviour of ICA by incorporating an additional minimization objective into the loss function [2]. It shows promising results in recovering the initial observation and provides an interesting discussion on the most suitable loading path to emphasize independence. The extension to the nonlinear case remains an open question due to non-identifiability issues. One interesting approach is to introduce an additional assumption at the level of the mixing function to recover identifiability. For example, this could involve assuming that the Jacobian components of the mixing function are independent in a non-statistical sense [3].