Test PineTree on Quijote latin hypercube rho -> halo

[maho3]

We want to test PineTree to constrain the halo biasing problem across the Quijote latin hypercube. It has currently only been tested on a single cosmology.

PineTree is a Neural Physical Engine, i.e. a sparse convolutional NN with few parameters which can be constrained with HMC.

Definition of done

• [ ] Plot of the 1-pt and 2-pt stats of a trained PineTree model applied to the Quijote LH, compared to the true halo field

[maho3]

Tagging @asiantaco

[maho3]

From Simon:

• I'm running some tests now with a very small CNN if in combination with Robin and the Poisson loss we can get a reasonable power spectrum
• I started to adapt my PineTree code to train on the latin hypercube. Should have results by next week.

[maho3]

From Simon:

The main update on the PineTree project is that I have set up to run the optimisation on the high resoltution density fields to see if the reconstruction for lower halo masses is better and also another run to go from fiducial density fields to the high resolution halos. Also I'm still testing the minimal training required for the PineTree model to see how feasible it is to train it on the whole LH simulations

[maho3]

PineTree paper here: https://arxiv.org/abs/2407.01391

This still requires tests of cosmology-dependence