source ~/.bashrc conda activate mace-scf mpirun -np 16 python /home/jhm/software/mace-tools/scripts/will_train.py \ --name="mace_scf_spice_2_low_force_filter_r6_128ch_charges_only" \ --train_file="../processed_maceoff_scf_spice_2/train" \ --test_file "../processed_maceoff_scf_spice_2/test" \ --valid_file="../processed_maceoff_scf_spice_2/val" \ --statistics_file="../processed_maceoff_scf_spice_2/statistics.json" \ --config_type_weights='{"Default":1.0}' \ --model="Polarizable" \ --hidden_irreps='128x0e + 128x1o' \ --loss="energy_forces_atomic_multipoles" \ --distributed \ --num_workers=4\ --num_interactions=2 \ --eval_interval=2 \ --batch_size=32 \ --valid_batch_size=32 \ --max_num_epochs=200\ --error_table="DensityEnergyRMSE" \ --swa\ --lr=5e-2 \ --swa_lr=5e-3 \ --start_swa=100\ --device="cuda"\ --kspace_cutoff_factor=1.0 \ --field_dependence_type="biased_local_linear" \ --energy_key="energy" \ --forces_key="forces" \ --atomic_multipoles_weight=10 \ --swa_atomic_multipoles_weight=1000 \ --atomic_multipoles_max_l=0 \ --atomic_multipoles_smearing_width=1.5 \ --keep_checkpoints \ --atomic_multipoles_key="mbis_multipoles" \ --default_dtype="float64" \ --restart_latest \ --scf_training_options='{"num_scf_steps":0, "constant_charge":True, "mixing_parameter": 1.0, "use_autograd_forces": True}'