Energy hyperparameter has an offset?

[bastonero]

Dear FLARE Developers,
I am using FLARE and it is amazing! Thanks a lot for this precious work!
Currently I am trying to train some snapshots of some materials ("offline" mode, using SGP Wrapper + B2 descriptors; see also attachment). Nevertheless, for different cases, MAE on forces and stresses are fine (~0.3 eV/Ang MAE for forces), whereas the energies have always a bigger error.
Inspecting the output, I noticed that the hyperparameter of the energy is ~, and if I subtract ( MAV-S_E ) / Nat ~ 0.1 eV, which would be a reasonable error. Here is the last part of the sdtout

Mean absolute errors & Mean absolute values
energy mae: 28662.2233 eV
energy mav: 28867.0624 eV
stress mae: 0.0021 eV/A^3
stress mav: 0.0051 eV/A^3
forces mae: 0.3296 eV/A
forces mav: 0.9431 eV/A
type 82 forces mae: 0.3919 eV/A
type 82 forces mav: 0.1833 eV/A
type 53 forces mae: 0.2898 eV/A
type 53 forces mav: 0.8769 eV/A
type 55 forces mae: 0.3865 eV/A
type 55 forces mav: 1.9017 eV/A

Uncertainty: [0.00388083 0.         0.        ]
Time of Update GP: 1.90 s
Time of Train Hyps: 98.09 s

GP hyperparameters:
Hyp0 : Hyp0                           = -2.1886
Hyp1 : Hyp1                           = 10.3087
Hyp2 : Hyp2                           = -28771.5235
Hyp3 : Hyp3                           = 0.5335
Hyp4 : Hyp4                           = 0.0021
Likelihood: -6066.0869
Likelihood gradient: [ 6.59480700e+01  4.52025492e+02  9.21402761e-06 -3.87215656e+02
  2.15562856e+03]

Any suggestion how to fix this?

Here is the input.yaml:

# ===================== STRUCTURE ======================= #
supercell: 
  file: dataset.xyz
  format: extxyz
  index: 0
# ==================== DFT CALCULATOR ===================== #
dft_calc:
  name: FakeDFT
  kwargs: {} 
  params: {}
# ======================= GP/ML ========================= #
flare_calc:
  gp: SGP_Wrapper
  kernels:
    - name: NormalizedDotProduct
      sigma: 2.0
      power: 2
    - name: NormalizedDotProduct
      sigma: 2.0
      power: 0.5
  descriptors:
    - name: B2
      nmax: 16
      lmax: 8
      cutoff_function: quadratic
      radial_basis: chebyshev
      #cutoff_matrix: [[5.0, 4.0], [4.0, 5.0]]
    - name: B2
      nmax: 16
      lmax: 8
      cutoff_function: quadratic
      radial_basis: chebyshev
  energy_noise: 0.04
  forces_noise: 0.05
  stress_noise: 0.005
  species:
    - 53 # I                                            
    - 55 # Cs
    - 82 # Pb 
  single_atom_energies: # ['Cs', -632.027637488, 'I', -2702.7083376388rm, 'Pb', -11689.338427714]
  #  - 2702.7083376388
  #  - 632.027637488
  #  - 11689.338427714
   - 0
   - 0
   - 0
  cutoff: 5.0
  variance_type: local
  max_iterations: 1000
  use_mapping: False
  energy_training: true
  force_training: true
  stress_training: true
# ======================= OTF & MD ========================= #
otf: 
  mode: fresh
  md_engine: Fake
  md_kwargs: 
      filenames:
        - dataset.xyz
      format: extxyz
      index: ":"
      io_kwargs: {}                            
  dt: 0.001                         
  number_of_steps: 84 # <= num. frames          
  output_name: offline            
  init_atoms: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]        
  std_tolerance_factor: -0.01
  max_atoms_added: -1
  train_hyps: [0, 84]
  write_model: 1
  update_style: threshold
  build_mode: bayesian
  update_threshold: 0.001
  force_only: False

[YuuuXie]

Hi @bastonero I hope you have solved this problem already, but just figured out there exists such a question and posting my late answer here. Yes, the energy noise can be optimized to a very large value, which is trapped in a local minimum, and this is mostly due to the very few energy labels, compared to force labels. You can try starting with different initializations of hyperparamters, or you can set the train_hyps such that there is no training of the hyper parameters such that the model will keep using your initial hyper parameters. In many cases, the model performance is not super sensitive to hyper parameters.