4.12. Fit electronic density of states (DOS) 

Note

Supported backends: TensorFlow

Here we present an API to DeepDOS model, which can be used to fit electronic density of state (DOS) (which is a vector).

See the PRB paper for details.

In this example, we will show you how to train a model to fit a silicon system. A complete training input script of the examples can be found in

$deepmd_source_dir/examples/dos/input.json

The training and validation data are also provided our examples. But note that the data provided along with the examples are of limited amount, and should not be used to train a production model.

Similar to the input.json used in ener mode, training JSON is also divided into model, learning_rate, loss and training. Most keywords remain the same as ener mode, and their meaning can be found here. To fit the dos, one needs to modify model/fitting_net and loss.

4.12.1. The fitting Network

The fitting_net section tells DP which fitting net to use.

The JSON of dos type should be provided like

	"fitting_net" : {
		"type": "dos",
		"numb_dos": 250,
		"sel_type": [0],
		"neuron": [120,120,120],
		"resnet_dt": true,
		"fparam": 0,
		"seed": 1,
	},

type specifies which type of fitting net should be used. It should be dos.
numb_dos specifies the length of output vector (density of states), which the same as the NEDOS set in VASP software, this argument defines the output length of the neural network. We note that the length of dos provided in training set should be the same.
The rest arguments have the same meaning as they do in ener mode.

4.12.2. Loss

DeepDOS supports trainings of the global system (a global dos label is provided in a frame) or atomic system (atomic labels atom_dos is provided for each atom in a frame). In a global system, each frame has just one dos label. For example, when fitting dos, each frame will just provide a 1 x numb_dos vector which gives the total electronic density of states. By contrast, in an atomic system, each atom in has a atom_dos label. For example, when fitting the site-projected electronic density of states, each frame will provide a natom x numb_dos matrices,

The loss section tells DP the weight of these two kinds of loss, i.e.

loss = pref * global_loss + pref_atomic * atomic_loss

The loss section should be provided like

	"loss" : {
		"type": "dos",
		"start_pref_dos": 0.0,
		"limit_pref_dos": 0.0,
		"start_pref_cdf": 0.0,
		"limit_pref_cdf": 0.0,
		"start_pref_ados": 1.0,
		"limit_pref_ados": 1.0,
		"start_pref_acdf": 0.0,
		"limit_pref_acdf": 0.0
	},

type should be written as dos as a distinction from ener mode.
pref_dos and pref_ados, respectively specify the weight of global and atomic loss. If set to 0, the corresponding label will not be included in the training process.
We also provides a combination training of vector and its cumulative distribution function cdf, which can be defined as

\[D(\epsilon) = \int_{e_{min}}^{\epsilon} g(\epsilon')d\epsilon'\]

4.12.3. Training Data Preparation

The global label should be named dos.npy/raw, while the atomic label should be named atomic_dos.npy/raw. If wrongly named, DP will report an error.

To prepare the data, we recommend shifting the DOS data by the Fermi level.

4.12.4. Train the Model

The training command is the same as ener mode, i.e.

dp train input.json

The detailed loss can be found in lcurve.out:

#  step      rmse_trn   rmse_ados_trn   rmse_ados_lr
    1.11e+00      1.11e+00    1.0e-03
    5.00e-02      5.00e-02    1.0e-03
    4.70e-02      4.70e-02    1.0e-03
    6.45e-02      6.45e-02    1.0e-03
    3.39e-02      3.39e-02    1.0e-03
    4.60e-02      4.60e-02    1.0e-03
    3.98e-02      3.98e-02    1.0e-03
    9.50e-02      9.50e-02    1.0e-03
    5.49e-02      5.49e-02    1.0e-03
    5.57e-02      5.57e-02    1.0e-03
    3.73e-02      3.73e-02    1.0e-03
    4.33e-02      4.33e-02    1.0e-03
    3.27e-02      3.27e-02    1.0e-03
    3.68e-02      3.68e-02    1.0e-03
    3.09e-02      3.09e-02    1.0e-03
    3.42e-02      3.42e-02    1.0e-03
    5.62e-02      5.62e-02    1.0e-03
    6.12e-02      6.12e-02    1.0e-03
    4.10e-02      4.10e-02    1.0e-03
    5.30e-02      5.30e-02    1.0e-03
    3.85e-02      3.85e-02    1.0e-03

4.12.5. Test the Model

In this earlier version, we can use dp test to infer the electronic density of state for given frames.

$DP freeze -o frozen_model.pb

$DP test -m frozen_model.pb -s ../data/111/$k -d ${output_prefix} -a -n 100

if dp test -d ${output_prefix} -a is specified, the predicted DOS and atomic DOS for each frame is output in the working directory

${output_prefix}.ados.out.0   ${output_prefix}.ados.out.1  ${output_prefix}.ados.out.2  ${output_prefix}.ados.out.3
${output_prefix}.dos.out.0   ${output_prefix}.dos.out.1  ${output_prefix}.dos.out.2  ${output_prefix}.dos.out.3

for *.dos.out.*, it contains matrix with shape of (2, numb_dos), for *.ados.out.*, it contains matrix with shape of (2, natom x numb_dos),

# frame - 0: data_dos pred_dos
000000000000000000e+00 1.963193264917645342e-03
000000000000000000e+00 1.178440836781313727e-03
000000000000000000e+00 1.441258071790407769e-04
000000000000000000e+00 1.787297933314058174e-03
000000000000000000e+00 1.901603280243024940e-03
000000000000000000e+00 2.279848925571981155e-03
000000000000000000e+00 2.149355854688561607e-03
000000000000000000e+00 1.829848459515726056e-03
000000000000000000e+00 1.905156512419792225e-03