Optimizing ground state of frustrated lattices?

[jeohjee]

I have compared the results of the G-CNN honeycomb tutorial to the exact diagonalization result and they seem to match very well. However, when I try to do a similar a G-CNN ground state optimization for a extended Heisenberg triangular lattice (J1 >0, J2=0 or J2>0), I consistently find a result that is far away from the ground state energy of the ED. I tried this for a small 8 site and 16 site triangular lattices both with very small J2=0.01 J1 and larger J2 = 0.25J1.

To solve this problem, I have tried to play with the parameter "equal_amplitudes" as in the tutorial and in a paper [PRR 2, 033075 (2020)] it has been mentioned that for difficult lattice models one should first optimize the complex phases of the wavefunction by keeping the amplitudes constant and only after that optimize also the amplitudes. This has not helped me (I actually don't see any change in the variational ground state energy when running the VMC with the equal_amplitudes=True) . I also tried BRM and symmetrized BRM models without luck.

Are there any other tricks that I could use?

[jeohjee]

Never mind, I was missing a setting param_dtype = jnp.complex128 in my model. Now I am able to obtain ground state energies that are close to the ED result.