nep.in reference

nep.in sets the model architecture and the training hyperparameters, one keyword per line (# starts a comment). The architecture keywords follow GPUMD, so a GPUMD nep.in describes the same model once its GPUMD-only lines (lambda_1, population, generation, …) are removed. Only the keywords listed on this page are accepted: any other one — a typo included — stops the run with an error that names it and its line.

nep.in

type       3 Cr Co Ni
version    4
zbl        2.5
use_typewise_cutoff_zbl 0.7
cutoff     6 4
n_max      8 8
basis_size 12 12
l_max      4 2 0
neuron     80

epoch            600
batch            16
lr               5e-3
scheduler_factor 0.5
stage2           1
start_stage2     300
stage2_lr        5e-4
stage2_lambda_e  2.0

Model architecture

Keyword	Default	Description
type	required	N name1 name2 ... — number and names of the element types.
version	4	NEP version; only NEP4 is implemented.
cutoff	8 4	Radial and angular cutoff in Å. Per species, as in GPUMD: cutoff rR1 rA1 rR2 rA2 …, one radial/angular pair per element in type order; an element pair uses the mean of the two values.
n_max	6 6	Radial and angular expansion orders.
basis_size	6 6	Chebyshev basis size, radial and angular.
l_max	4 1 0	L_3b q_222 q_1111 q_112 q_123 q_233 q_134: the largest L of the 3-body terms (1–8), then up to six flags that switch on each higher-body invariant (GPUMD order).
neuron	30	Neurons in the hidden layer.
zbl	off	ZBL outer cutoff in Å; switches on the short-range repulsion. A file name instead of a number (zbl zbl.in) reads GPUMD's flexible ZBL table.
use_typewise_cutoff_zbl	off	use_typewise_cutoff_zbl <factor>: per-pair ZBL outer cutoff = min(factor × (Ri + Rj), zbl) with covalent radii R, inner cutoff 0. The factor is required (0.7 recommended, ≥ 0.5).

Cutoff rules

Checked when nep.in is read:

• the angular cutoff is ≥ 3 Å and ≤ the radial cutoff, for every species;
• the radial cutoff is ≤ 100 Å;
• the ZBL outer cutoff (zbl, and every row of zbl.in) is between 1 and 3 Å, so ZBL always lies inside the angular neighbor list.

Flexible ZBL

zbl zbl.in reads one line per element pair, in the order 1-1, 1-2, …, 1-n, 2-2, …, n-n:

zbl.in

rc_inner rc_outer a1 a2 a3 a4 a5 a6 a7 a8

To change only the cutoffs, keep the universal coefficients 0.18175 3.1998 0.50986 0.94229 0.28022 0.4029 0.02817 0.20162. The table is stored in nep.txt, so GPUMD, NEP_CPU and LAMMPS use it without the file; use_typewise_cutoff_zbl is ignored.

Training hyperparameters

Keyword	Default	Description
epoch	600	Total training epochs.
batch	32	Structures per gradient step (per GPU with train_nep_sharded).
lr	0.01	Initial learning rate.
stop_lr	1e-6	Lower bound of the learning rate.
lambda_e	0.01	Energy loss weight.
lambda_f	1.0	Force loss weight.
lambda_v	0.01	Virial loss weight.
weight_decay	1e-4	AdamW decoupled weight decay on all trainable parameters; 0 uses plain Adam.
max_grad_norm	10.0	Gradient clipping threshold.
lr_scheduler	plateau	plateau (reduce on plateau) or step (fixed interval); both stages use this mode.
scheduler_patience	15	plateau: epochs without improvement before a reduction. step: epochs between reductions.
scheduler_factor	0.7	Factor applied at each reduction.
early_stop	0	Stop when the monitored loss has not improved for N epochs (0 = off). Per stage: a stage-1 plateau moves on to stage 2. Use a value larger than scheduler_patience.
stage2	0	1 switches on the second, energy-focused stage.
start_stage2	half of epoch	Epoch at which stage 2 starts.
stage2_lr	1e-3	Learning rate at the start of stage 2.
stage2_lambda_e	1.0	Stage 2 energy weight.
stage2_lambda_f	0.05	Stage 2 force weight.
stage2_lambda_v	0.1	Stage 2 virial weight.
stage2_scheduler_patience	scheduler_patience	Stage 2 scheduler patience.
stage2_scheduler_factor	scheduler_factor	Stage 2 reduction factor.

The monitored loss is the validation loss when a validation set is used, otherwise the training loss. Options that are not hyperparameter values — device, precision, validation data, checkpoints — are arguments of the Python function; see Training.