On the possible environmental causes of inconsistencies between the ground state energies of water molecules and the results reported in the article

[KingDamLiu]

We found that when calculating the ground - state energy of molecules using ferminet, the results are inconsistent with the energies reported in the literature. Such discrepancies are observed in different versions of the code and different operating environments. Below, we conducted relevant tests using the water molecule as an example.
The calculation code is as follows:

from ferminet.configs.excited.oscillator import get_config
from ferminet import base_config
from ferminet import train
from absl import logging
import sys
 
logging.get_absl_handler().python_handler.stream = sys.stdout
logging.set_verbosity(logging.INFO)
cfg = get_config()
cfg.system.molecule_name = 'H2O'
cfg.system.states = 0
# cfg.optim.laplacian = 'folx'
cfg.network.network_type = f'psiformer'
cfg = base_config.resolve(cfg)
train.train(cfg)

Code version details:
Version 1: 266c025
Version 2: c6f9db5
Operating environment details:
Environment 1: jax 0.5.0, kfac 0.0.6
Environment 2: jax 0.5.3, kfac 0.0.6
Under Version 1, we conducted tests using Environment 1 and Environment 2 respectively. The tests were completed using psiformer without folx. The result in Environment 1 was approximately 76.428 Hatree, and its energy convergence curve was relatively stable. In contrast, the result in Environment 2 was approximately 76.510 Hatree, with a highly fluctuating energy convergence curve.
Under Version 2, we only conducted tests in Environment 2. The results were all around 76.510 Hatree, and the energy convergence curves also fluctuated significantly. Four types of tests were carried out, including ferminet and psiformer, with and without folx, and the results were consistent.
Previously, we also conducted experiments using other versions of the code in the jax 0.4.24 and jax 0.4.34 environments, and no convergence stability issues were found. However, regarding the ground - state energy of the water molecule, except for obtaining 76.4380 Hatree (consistent with the result reported in the article) using psiformer in the jax 0.4.24 environment, the results in other cases were 76.428 Hatree and 76.434 Hatree. In the same experimental environment as the water molecule reproduction, the CO molecule failed to reproduce the result of 113.324 Hatree reported in the article; the actual result was 113.319 Hatree. However, the results for the smaller BH molecule could almost always be reproduced.
Based on all the test results, we preliminarily speculate that the differences in the calculation results may be caused by different operating environments. In particular, the latest environment may lead to convergence stability issues.
If it is convenient, we hope to obtain information about the operating environment used by the authors in the calculation of the paper "Accurate computation of quantum excited states with neural networks".

[jsspencer]

I have not been able to reproduce the discrepancy on the BH system you reported in #96. Am I correct in understanding that you now reproduce the reported energies for BH? If so, what changed?

[KingDamLiu]

Yes, in other environments, I can easily reproduce the reported BH energy of 28.289. However, on larger systems such as H2O and CO, it is difficult to reproduce the results reported in the literature. The above results can always be repeatedly obtained in their respective environments and code versions.