Format of sannp.prop

Executing sannp --temp outputs the template of sannp.prop.

A Line that initial character is # or ! is dealt with as a comment.

restart

default:

0

Start learning from the beginning in case of 0. Read the information of the neural network from sannp.data, sannp.data_e or sannp.data_q to resume the learning in case of 1.

insituTest

default:

0

Perform in-situ tests to calculate and output RSME against the test data in each epoch during learning in case of 1. Not perform in-situ tests in case of 0.

withCharge

default:

0

Calculate charges (1) or not (0).

withHDNNP

default:

1

Use HDNNP method (1) of setting the total energy in the system as training data or SANNP method (0) of setting energy divided into each atom as training data.

withLJlike

default:

0

Use Δ-NNP, combination of LJ-like force field and neural network force field (1) or not(0). Can not be used with Δ-NNP using ReaxFF at the same time. withClassical 1 is the same as withLJlike 1.

withReaxFF

default:

0

Use Δ-NNP, combination of ReaxFF and neural network force field (1) or not(0). For ReaxFF, parameter definition file ffield.reax is required. Can not be used with Δ-NNP using LJ-like force field at the same time. withClassical 2 is the same as withReaxFF 1.

rcutReaxFF

default:

5.0

For Δ-NNP using ReaxFF, specify the cutoff radius(Å) of ReaxFF.

rateReaxFF

default:

0.5

For Δ-NNP using ReaxFF, specify the contribution (mixing rate) of ReaxFF when calculating energy and force.

directSF

default:

-1

Normalize symmetry functions in each mini batch (1) or in whole sample (0); in case of a negative value, normalize in whole sample for Behler symmetry function, in mini batch for Many-Body symmetry function.

regularElem

default:

1

When characters follow element names in the training data (e.g., Fe1, Fe2), specify whether to treat them as the same element (1) or as different elements (0) during training.

maxForce

default:

10.0

Specify the threshold (eV/Å) to eliminate outliers with too large force in training data. Not eliminate in case of 0 or lower.

minEDev

default:

0.5

Specify the lower limit (eV) of variance for normalizing of atomic energy.

maxEDev

default:

10

Specify the upper limit (eV) of variance for normalizing of atomic energy.

minQDev

default:

0.1

Specify the lower limit (e) of variance for normalizing of atomic charges.

maxQDev

default:

10

Specify the upper limit (e) of variance for normalizing of atomic charges.

symmFunc

default:

chebyshev

Specify a symmetry function. behler, chebyshev and many-body are available.

elemWeight

default:

1

Use weighted symmetry functions (1) or not (0). Available only when Behler symmetry function or Chebyshev symmetry function is used.

tanhCutoff

default:

1

Use functions composed of tanh (1) or cos (0) as cutoff functions .

m2

default:

100

Specify the parameter M₂ of Many-Body symmetry functions.

m3

default:

10

Specify the parameter M₃ of Many-Body symmetry functions.

rinner

default:

0.0

Specify the parameter R_inner (Å) of Many-Body functions.

router

default:

6.0

Specify the parameter R_outer (Å) of Many-Body functions.

numRadius

default:

50

Specify the number of radial components of Chebyshev symmetry functions.

numAngle

default:

30

Specify the number of angular components of Chebyshev symmetry functions.

rcutRadius

default:

6.0

Specify the cutoff radius R_c (Å) of radial components of Chebyshev symmetry functions.

rcutAngle

default:

6.0

Specify the cutoff radius R_c (Å) of angular components of Chebyshev symmetry functions.

models

default:

16

Create specified number of neural network models and perform learning in parallel. When used as the force field, the average value of energy/force output from each neural network becomes the output of the force field. If 1 is specified, it results in the ordinary method in which the force field is defined by one neural network model.

layers

default:

2

Specify the number of hidden layers of a neural network.

nodes

default:

40

Specify the number of nodes of of a neural network.

activ

default:

twtanh

Specify an activation function of a neural network. asis (not use), sigmoid, tanh, twtanh (twisted tanh), eLU, and GELU are available.

lbfgs

default:

32

Specify an optimization algorithm at learning. In case of 0, use Adam method. In case of 1 or larger, use L-BFGS method with the specified value as the number of histories.

lineSearch

default:

more-thuente

Specify a linear search algorithm used in L-BFGS method. more-thuente, armijo, wolfe, and srong-wolfe are available.

lineSteps

default:

32

Specify the maximum value of the number of trials of linear search used in L-BFGS method.

batchs

default:

0

Specify the mini batch size at learning. In case of 0 or lower, not use mini batches but use a full batch (whole sample) for learning.

epochs

default:

500

Specify the upper limit of the number of repeats (epoch) at learning.

epochsStore

default:

1000

Specify the interval to save data of neural network into a file at learning. In case of 0 or lower, store only at the end.

epochsOnlyE

default:

250

Use only energy for learning until the nuber of repeats (epoch) at learning becomes larger than this number. After that, use both energy and force for learning.

epochsApproxF

default:

500

Use approximate differentiation with Double Backward method when calculating from energy an error of force acting on an atom until the number of repeats (epoch) becomes larger than this number. After that, strictly calculate a differentiate.

superEpochs

default:

0

Setting of Super Epoch method, in which the training data is split in subsets and serially perform the training using each subset.

Specify the number of subsets (= number of Super Epochs). In case of 0 or lower, the value is automatically set based on geomsEpoch.

geomsEpoch

default:

500 (2500 for GPU version sannp_gpu)

Specify the number of training data per Super Epoch. In case of 0 or lower, or larger than the training data number, it is automatically set to the same value as the training data number.

sqrtLoss

default:

0

Specify the scale factor of a loss function with absolute value (1) or squared value (0).

renormLoss

default:

0

Specify whether a loss function should be normalized (1) or not (0).

approxForce

default:

0

Select switching approximate and strict differentiation with epochsApproxF (0) or using always approximate one (1) when calculating from energy an error of force acting on an atom.

rmseEnergy

default:

0.01

Specify the threshold (eV/atom) of residual (RMS) of energy to determine whether learning has converged or not.

rmseForce

default:

0.10

Specify the threshold (eV/Å) of residual (RMS) of force to determine whether learning has converged or not.

rmseCharge

default:

0.01

Specify the threshold (e) of residual (RMS) of charge to determine whether learning has converged or not.

coefEnergy

default:

1.00

Specify the scale factor (1/eV or 1/eV²) of a loss function of energy.

coefForce

default:

1.00

Specify the scale factor (Å/eV or (Å/eV)²) of a loss function of force.

coefCharge

default:

1.00

Specify the scale factor (1/e or 1/e²) of a loss function of charge.

learnRate

default:

1.0e-4

Specify the initial value of learning rate.

learnRateFinal

default:

1.0e-4

Specify the lower limit of learning rate.

learnRateDecay

default:

0.9999

Specify the attenuation factor of learning rate.

adamBeta1

default:

0.9

Specify the hyper parameter (β₁ of Adam method) at learning.

adamBeta2

default:

0.999

Specify the hyper parameter (β₂ of Adam method) at learning.

classicalTry

default:

64

Specify the upper limit of the number of repeats when optimizing a classical force field used in Δ-NNP.

classicalLower

default:

-50.0

When optimizing a classical force field used in Δ-NNP, penalty functions will be applied so that lower energy than this value (eV) has seldomly appeared.

gpuThreads

default:

256

(GPU version) Number of threads per block in CUDA. Upper limit is 1024 (specification of CUDA). Multiple of 32 (warp size) is appropriate.

This setting is about the threads on GPU, and not related to thread parallel (OpenMP parallel) on CPU.

gpuAtomBlock

default:

512

(GPU version) Number of atoms to be processed at a time when calculating the symmetry function using GPU.

endProperty

Subsequent file contents are treated as comments.