Probing Physics-Informed Generative Representations of Many-Body Wavefunctions

The inclusion of physics-informed constraints has been shown to improve the performance of neural networks, particularly within the context of solving partial differential equations. Here, we adopt a similar strategy for generative models, including GAN and VAE, and apply this method to simulated ground state snapshots of 1D XXZ spin chains. We investigate how incorporating relevant observables into the models helps them recover the underlying physical structure of the input data, including ferromagnetic, antiferromagnetic, and paramagnetic phases. By choosing the optimal physical features, this method can be used in order to more accurately represent unknown wavefunctions when given experimental data.