Large initial energy RMSE when finetuning

[tomdemeyere]

I started to finetune the SevenNet-l3i5 model with a dataset containing ~ 3000 structures of slabs and nanoparticles. The original energy RMSE is extremely high (2500 eV/atom) while forces are ok (400 meV/Å). Is there something I don't understand with energy references?

Here is my finetune.yaml:

model:
  act_gate:
    e: silu
    o: tanh
  act_radial: silu
  act_scalar:
    e: silu
    o: tanh
  channel: 128
  chemical_species: auto
  cuequivariance_config: {}
  cutoff: 5.0
  cutoff_function:
    cutoff_function_name: poly_cut
    poly_cut_p_value: 6
  interaction_type: nequip
  irreps_manual:
  - 128x0e
  - 128x0e+64x1e+32x2e+32x3e
  - 128x0e+64x1e+32x2e+32x3e
  - 128x0e+64x1e+32x2e+32x3e
  - 128x0e+64x1e+32x2e+32x3e
  - 128x0e
  is_parity: false
  lmax: 3
  lmax_edge: -1
  lmax_node: -1
  num_convolution_layer: 5
  radial_basis:
    bessel_basis_num: 8
    radial_basis_name: bessel
  readout_as_fcn: false
  self_connection_type: linear
  train_denominator: false
  train_shift_scale: true
  train_avg_num_neigh : true 
  use_bias_in_linear: false
  use_modal_node_embedding: false
  use_modal_output_block: false
  use_modal_self_inter_intro: false
  use_modal_self_inter_outro: false
  weight_nn_hidden_neurons:
  - 64
  - 64

train:
    random_seed: 1
    is_train_stress: False
    epoch: 100

    force_loss_weight : 100.0
    stress_loss_weight: 0.0
    loss: huber
    loss_param:
        delta: 0.01

    optimizer: adamw
    optim_param:
        lr: 0.0002
    scheduler: 'exponentiallr'
    scheduler_param:
        gamma: 0.99

    csv_log: log.csv

    error_record:
        - ['Energy', 'RMSE']
        - ['Force', 'RMSE']
        - ['Stress', 'RMSE']
        - ['TotalLoss', 'None']

    loss_param: {}

    #scheduler: onecyclelr
    #scheduler_param:
    #    anneal_strategy: cos
    #    div_factor: 25.0
    #    epochs: 2
    #    final_div_factor: 10000.0
    #    max_lr: 0.0002
    #    pct_start: 0.005
    #    total_steps: 1007534
    #stress_loss_weight: 1.0e-06
    train_shuffle: true
    use_weight: false

    per_epoch: 10

    continue:
        reset_optimizer: True
        reset_scheduler: True
        reset_epoch: True
        checkpoint: checkpoint_l3i5.pth
        use_statistic_values_of_checkpoint: False

data:
  batch_size: 8
  compute_statistics: true
  #use_species_wise_shift_scale: True
  shift: 'elemwise_reference_energies'
  data_format: ase
  data_format_args: {}
  load_trainset_path: training_set.xyz
  load_validset_path: validation_set.xyz
  preprocess_num_cores: 1
  use_modal_wise_scale: false
  use_modal_wise_shift: false

[YutackPark]

Have you drawn parity plot using 7net-l3i5 at your training set? I see two possibilities here 1. Some outlier in the training data 2. Huge energy shift If the cause is 1. Check the outlier data whether DFT calculation is not converted or it has too high energy that physically meaningless In the case of 2, which functional or pseudopotential do you use? 7net l3i5 trained on PBE52 You can get data for the parity plot from `sevenn_inference 7net-l3i5 {path to your training dataset}`

…

----------------------------------------------------------------- *Yutack Park* Ph.D. Candidate Materials Data and Informatics Lab. Department of Materials Science and Engineering Seoul National University *E-mail:* ***@***.*** ***@***.***> *Homepage:* http://mtcg.snu.ac.kr ----------------------------------------------------------------- 2025년 3월 21일 (금) 오전 6:16, Tom Demeyere ***@***.***>님이 작성:

I started to finetune the SevenNet-l3i5 model with a dataset containing ~ 3000 structures of slabs and nanoparticles. The original energy RMSE is extremely high (2500 eV/atom) while forces are ok (400 meV/Å). Is there something I don't understand with energy references? Here is my finetune.yaml: model: act_gate: e: silu o: tanh act_radial: silu act_scalar: e: silu o: tanh channel: 128 chemical_species: auto cuequivariance_config: {} cutoff: 5.0 cutoff_function: cutoff_function_name: poly_cut poly_cut_p_value: 6 interaction_type: nequip irreps_manual: - 128x0e - 128x0e+64x1e+32x2e+32x3e - 128x0e+64x1e+32x2e+32x3e - 128x0e+64x1e+32x2e+32x3e - 128x0e+64x1e+32x2e+32x3e - 128x0e is_parity: false lmax: 3 lmax_edge: -1 lmax_node: -1 num_convolution_layer: 5 radial_basis: bessel_basis_num: 8 radial_basis_name: bessel readout_as_fcn: false self_connection_type: linear train_denominator: false train_shift_scale: true train_avg_num_neigh : true use_bias_in_linear: false use_modal_node_embedding: false use_modal_output_block: false use_modal_self_inter_intro: false use_modal_self_inter_outro: false weight_nn_hidden_neurons: - 64 - 64 train: random_seed: 1 is_train_stress: False epoch: 100 force_loss_weight : 100.0 stress_loss_weight: 0.0 loss: huber loss_param: delta: 0.01 optimizer: adamw optim_param: lr: 0.0002 scheduler: 'exponentiallr' scheduler_param: gamma: 0.99 csv_log: log.csv error_record: - ['Energy', 'RMSE'] - ['Force', 'RMSE'] - ['Stress', 'RMSE'] - ['TotalLoss', 'None'] loss_param: {} #scheduler: onecyclelr #scheduler_param: # anneal_strategy: cos # div_factor: 25.0 # epochs: 2 # final_div_factor: 10000.0 # max_lr: 0.0002 # pct_start: 0.005 # total_steps: 1007534 #stress_loss_weight: 1.0e-06 train_shuffle: true use_weight: false per_epoch: 10 continue: reset_optimizer: True reset_scheduler: True reset_epoch: True checkpoint: /iridisfs/scratch/ba3g18/IridisX/toms_playground/sevenn/smaller/checkpoint_l3i5.pth use_statistic_values_of_checkpoint: False data: batch_size: 8 compute_statistics: true #use_species_wise_shift_scale: True shift: 'elemwise_reference_energies' data_format: ase data_format_args: {} load_trainset_path: training_set_no_duplicates_final.xyz load_validset_path: validation_set.xyz preprocess_num_cores: 1 use_modal_wise_scale: false use_modal_wise_shift: false — Reply to this email directly, view it on GitHub <#191>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A2RQY2Y5SY7L6ALJACNALST2VMV35AVCNFSM6AAAAABZOPZBBGVHI2DSMVQWIX3LMV43ERDJONRXK43TNFXW4OZYGEYTCMRRGE> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>