10.1. ALM: Input files#

Format of input files#

Each input file should consist of entry fields. Available entry fields are

&general, &interaction, &cutoff, &cell, &position, and &optimize.

Each entry field starts from the key label &field and ends at the terminate character “/”. (This is equivalent to Fortran namelist.)

For example, &general entry field of program alm should be given like

&general
  # Comment line
  PREFIX = prefix
  MODE = optimize
/

Multiple entries can be put in a single line. Also, characters put on the right of sharp (“#”) are neglected. Therefore, the above example is equivalent to the following:

&general
  PREFIX = prefix; MODE = optimize  # Comment line
/

Each variable must be given inside the appropriate entry field.

List of supported input variables#

Description of input variables#

“&general”-field#

  • PREFIX-tag : Job prefix to be used for names of output files

Default:

None

Type:

String


  • MODE-tag = optimize | suggest

optimize

Estimate harmonic and anharmonic IFCs.
This mode requires an appropriate &optimize field.

suggest

Suggests the displacement patterns necessary
to estimate harmonic and anharmonic IFCs.
Default:

None

Type:

String


  • NAT-tag : Number of atoms in the supercell

Default:

The number of entries in the &position field (or in STRUCTURE_FILE)

Type:

Integer


  • KD-tag : List of atomic species

Default:

None

Type:

Array of strings

Example:

In the case of GaAs, it should be KD = Ga As.


  • STRUCTURE_FILE-tag : File containing the input structure

Default:

None

Type:

String

Description:

When this tag is given, alm reads the lattice vectors, atomic species, and atomic coordinates from the specified file in the VASP POSCAR format. In this case, the &cell and &position fields are ignored, and NAT and KD are inferred from the structure file.

The POSCAR file may use direct or Cartesian coordinates. A Selective dynamics line is accepted, but the selective-dynamics flags themselves are ignored.

Note

A POSCAR is always interpreted in Angstrom (the VASP convention), independent of the LENGTH_UNIT tag.


  • SUPERCELL-tag : Transformation matrix from the input cell to the supercell

Default:

1

Type:

Integer matrix

Description:

This tag specifies the target supercell as \((\boldsymbol{a}_{s}, \boldsymbol{b}_{s}, \boldsymbol{c}_{s})^{T} = M (\boldsymbol{a}_{\mathrm{in}}, \boldsymbol{b}_{\mathrm{in}}, \boldsymbol{c}_{\mathrm{in}})^{T}\). If the input structure is already the supercell used for the force calculations, keep the default value.

The matrix can be written as one, three, or nine entries. One entry n means n n n; three entries n1 n2 n3 mean a diagonal matrix; nine entries give the full matrix in row-major order. The determinant must be a nonzero integer.

Example:

SUPERCELL = 2 2 2 creates a 2x2x2 supercell from the input cell. SUPERCELL = 2 0 0 0 2 0 0 0 1 creates a 2x2x1 supercell.


  • PRIMCELL-tag : Transformation matrix from the input cell to the primitive cell

Default:

1

Type:

Matrix or Auto

Description:

This tag specifies the primitive cell as \((\boldsymbol{a}_{p}, \boldsymbol{b}_{p}, \boldsymbol{c}_{p})^{T} = M (\boldsymbol{a}_{\mathrm{in}}, \boldsymbol{b}_{\mathrm{in}}, \boldsymbol{c}_{\mathrm{in}})^{T}\). The matrix can be written as one, three, or nine entries, using the same convention as SUPERCELL. Fractional entries such as 1/2 are accepted. The determinant of the inverse matrix must be a nonzero integer.

PRIMCELL = Auto uses spglib to detect the primitive cell from the input structure. This is useful when the input structure is a conventional cell or a supercell and the primitive-cell transformation is not obvious.

Example:

PRIMCELL = 1/2 0 0 0 1/2 0 0 0 1/2 defines a primitive cell whose lattice vectors are one half of the corresponding input-cell lattice vectors.


  • TOLERANCE-tag : Tolerance for finding symmetry operations

Default:

1.0e-3

Type:

Double


  • TOL_CONST-tag : Tolerance used when reducing the constraints between force constants

Default:

1.0e-6

Type:

Double

Description:

Numerical tolerance used to detect linearly dependent constraints (the rank of the constraint matrix) when the constraints of ICONST and the symmetry relations are reduced. The default value is usually sufficient; see also the note on the input precision under ICONST.


  • SYMMETRIZE_STRUCTURE-tag = 0 | 1

0

Use the input structure as given.

1

Symmetrize the lattice vectors and atomic positions to the space group that spglib detects
within TOLERANCE. The structure is not standardized: the cell choice, atom order, and
orientation are kept, the lattice changes only by a symmetric (rotation-free) strain, and
the atoms are not shifted as a whole. The symmetrized structure is used for all later steps
and written to the output files.
Default:

0

Type:

Integer

Description:

Useful when the input coordinates are given with only a few digits, e.g. 0.3333333333 instead of 1/3. The changes of the lattice vectors and atomic positions are printed to the standard output. The input displacements in DFSET are not modified. Not available with NONCOLLINEAR = 1 or with PERIODIC = 0 in any direction.

The symmetrization is applied to the input cell, i.e., the structure given by &cell and &position or by STRUCTURE_FILE, right after it is read. The supercell (SUPERCELL) and the primitive cell (PRIMCELL, including PRIMCELL = Auto) are then built from the symmetrized input cell, so they are symmetric as well.

When SYMMETRIZE_STRUCTURE = 0, ALM checks how far the input structure deviates from the ideal structure of its space group and prints a warning recommending SYMMETRIZE_STRUCTURE = 1 if a lattice-vector component or an atomic position deviates by more than \(10^{-6}\) bohr.


  • PRINTSYM-tag = 0 | 1

0

Symmetry operations won’t be saved in “SYMM_INFO”

1

Symmetry operations will be saved in “SYMM_INFO”

Default:

0

type:

Integer


  • FCSYM_BASIS-tag = Cartesian | Lattice

Cartesian, C

Symmetry reduction of force constant is performed in the Cartesian basis

Lattice, L

Symmetry reduction of force constant is performed in the \(\boldsymbol{a}_1, \boldsymbol{a}_2, \boldsymbol{a}_3\) basis

Default:

Cartesian (changed from Lattice in version 2.0)

type:

String

Description:

The calculation results should not depend on the choice of FCSYM_BASIS when LMODEL = ols. For other regression methods (enet, adaptive LASSO), an optimal value of the L1_ALPHA changes when you change the FCSYM_BASIS option.

FCSYM_BASIS = Cartesian is the default because it is required for the rotational invariance constraints. FCSYM_BASIS = Lattice can be more robust when the input structure is given with only a few digits, e.g., hexagonal cells with \(\sqrt{3}/2\) written as 0.866025. In such cases, setting SYMMETRIZE_STRUCTURE = 1 is recommended instead of, or in addition to, FCSYM_BASIS = Lattice.

For more details about the symmetry reduction of force constants, please see here.

Important

When FCSYM_BASIS = Lattice, the basis of force constants saved in PREFIX.fcs becomes the \(\boldsymbol{a}_1, \boldsymbol{a}_2, \boldsymbol{a}_3\) basis. Hence, to compare the values of force constants saved in PREFIX.fcs, you will have to change their basis to the Cartesian basis manually. The basis of force constants saved in PREFIX.xml is Cartesian irrespective of the FCSYM_BASIS value.

Also, imposing the constraints for rotational invariance with FCSYM_BASIS = Lattice is not supported. Therefore, if you want to apply the constraints for rotational invariance, please use FCSYM_BASIS = Cartesian.


  • MAGMOM-tag : List of magnetic moments

Default:

0 … 0 (NAT entries when NONCOLLINEAR = 0, 3xNAT entries when NONCOLLINEAR = 1.)

type:

Array of double

Example:

When a supercell contains 64 atoms and the local magnetic moments of the first 32 atoms are up and those of the last 32 atoms are down, please set the MAGMOM tag as MAGMOM = 32*1 32*-1. The wildcard (*) is available when NONCOLLINEAR = 0. For the noncollinear case (NONCOLLINEAR = 1), the wildcard is not supported. So, please give the magnetic moment explicitly as MAGMOM = 0 0 1 0 0 1 0 0 1 ... 0 0 -1 0 0 -1 ... (3\(\times\)NAT entries in one line).

Note

MAGMOM information is used only for generating space group operations. So, the values of the magnetic moment are somewhat arbitrary. For the 4\(\times\) 4\(\times\) 4 supercell of ferromagnetic bcc Fe (64 atoms), for instance, MAGMOM = 64*1 and MAGMOM = 64*2 give the same results. By contrast, MAGMOM = 32*1 32*2 of course gives a different result because it breaks the symmetry of the original lattice.


  • NONCOLLINEAR-tag = 0 | 1

Default:

0

type:

Integer

Description:

When NONCOLLINEAR = 1, the code accepts a noncollinear magnetic structure as an input to the MAGMOM tag and uses it for generating space group operations. The spin quantization axis is fixed to the (0,0,1) direction of the Cartesian axis.

Caution

Still experimental. Please use with care.


  • PERIODIC-tag = PERIODIC[1], PERIODIC[2], PERIODIC[3]

0

Do not consider periodic boundary conditions when
searching for interacting atoms.

1

Consider periodic boundary conditions when
searching for interacting atoms.
Default:

1 1 1

type:

Array of integers

Description:

This tag is useful for generating interacting atoms in low dimensional systems. When PERIODIC[i] is zero, periodic boundary condition is turned off along the direction of the lattice vector \(\boldsymbol{a}_{i}\).


  • NMAXSAVE-tag : The maximum order of anharmonic force constants written to the force-constant files (PREFIX.h5 and PREFIX.xml)

Default:

min(5, NORDER)

Type:

Integer

Example:

If your model includes anharmonic terms up to the sixth-order (NORDER = 5), but you want to avoid printing out the fifth-order and sixth-order IFCs, please set NMAXSAVE = 3.


  • FCS_ALAMODE-tag = 0 | 1

0

Write the force constants to PREFIX.h5 only

1

Additionally write the legacy PREFIX.xml file and the plain-text PREFIX.fcs listing

Default:

0

Type:

Integer

Description:

The HDF5 file PREFIX.h5 is the standard force-constant output and is always written. Set FCS_ALAMODE = 1 to also produce the legacy XML file and the human-readable PREFIX.fcs listing, e.g. for external tools that have not migrated to the HDF5 format yet. anphon reads both formats.


  • HESSIAN-tag = 0 | 1

0

Do not save the Hessian matrix

1

Save the entire Hessian matrix of the supercell as PREFIX.hessian.

Default:

0

type:

Integer


  • FC3_SHENGBTE-tag = 0 | 1

0

Do not save the third-order force constants for the ShengBTE code

1

Save the third-order force constants for the ShengBTE code in PREFIX.FORCE_CONSTANTS_3RD.
Default:

0

type:

Integer

Description:

The file has the layout of ShengBTE’s FORCE_CONSTANTS_3RD: the force constants in eV/Å3 and the lattice vectors of the periodic images in Å, independent of FCS_UNIT_OUTPUT. It is written only when NORDER >= 2; otherwise a warning is printed and no file is written. ShengBTE reads a file named FORCE_CONSTANTS_3RD, so remove the PREFIX. part or link the file. To convert existing force constants instead, see Converting force constants (phonopy, phono3py, ShengBTE).


  • FC4_SHENGBTE-tag = 0 | 1

0

Do not save the fourth-order force constants for the FourPhonon code

1

Save the fourth-order force constants for the FourPhonon code in PREFIX.FORCE_CONSTANTS_4TH.
Default:

0

type:

Integer

Description:

The file has the layout of FORCE_CONSTANTS_4TH read by FourPhonon, the four-phonon extension of ShengBTE: the force constants in eV/Å4 and the lattice vectors of the periodic images in Å, independent of FCS_UNIT_OUTPUT. It is written only when NORDER >= 3; otherwise a warning is printed and no file is written. Remove the PREFIX. part or link the file to FORCE_CONSTANTS_4TH.


  • FC_ZERO_THR-tag : Threshold value used when trimming force constants written to the force-constant files

Default:

1.0e-12

Type:

Double

Description:

FC_ZERO_THR defines the threshold of force constants to be saved in PREFIX.h5 (and PREFIX.xml). If the absolute value of a force constant is smaller than FC_ZERO_THR, it will NOT be printed out.

Note

If the harmonic force constants are calculated using a model potential (e.g., classical FF) where the interaction becomes zero beyond a certain cutoff radius, the default value of FC_ZERO_THR may drop small but nonzero force constants. This issue may be resolved by using a smaller FC_ZERO_THR, say FC_ZERO_THR = 1.0e-15. The force constants that become exactly zero due to symmetry and acoustic sum rule constraints will not be printed even when setting FC_ZERO_THR = 0.


  • FORMAT_PATTERN-tag = yaml | old

yaml

Write the displacement-pattern files in the YAML format

old

Write them in the plain-text format of ALAMODE 1.x

Default:

yaml

Type:

String

Description:

Format of the displacement-pattern files PREFIX.pattern_HARMONIC, PREFIX.pattern_ANHARM3, … written with MODE = suggest. displace.py --pattern_file reads both formats; old is kept for external scripts written for the 1.x format.


  • LENGTH_UNIT-tag = bohr | angstrom

Default:

bohr

Type:

String

Description:

Unit of the length-type input data: the &cell lattice vectors, the displacements in the DFSET file, and the &cutoff radii. The data are converted to Bohr internally, so the internal model and all outputs are unaffected. A STRUCTURE_FILE (POSCAR) is always interpreted in Angstrom, independent of this tag.


  • FORCE_UNIT-tag = Ry/bohr | eV/angstrom | Ha/bohr

Default:

Ry/bohr

Type:

String

Description:

Unit of the forces in the DFSET file, converted to Ry/bohr internally.


  • FCS_UNIT_OUTPUT-tag = Ry/bohr | eV/angstrom

Default:

Ry/bohr

Type:

String

Description:

Unit system of the force constants written to the HDF5 file (PREFIX.h5). With eV/angstrom, the force constant values (eV/angstromn), the lattice vectors, and the shift vectors (angstrom) are converted consistently, and the unit attributes stored in the file are updated accordingly; anphon reads these attributes and converts the data back internally. The XML (PREFIX.xml) and PREFIX.fcs outputs always stay in Rydberg atomic units.


“&interaction”-field#

  • NORDER-tag : The order of force constants to be calculated. Anharmonic terms up to \((m+1)\)th order will be considered with NORDER = \(m\).

Default:

None

Type:

Integer

Example:

NORDER = 1 to calculate harmonic terms only, NORDER = 2 to include cubic terms as well, and so on.


  • NBODY-tag : Entry for excluding multiple-body interactions from anharmonic force constants

Default:

NBODY = [2, 3, 4, …, NORDER + 1]

Type:

Array of integers

Description:

This tag may be useful for excluding multi-body clusters which are supposedly less important. For example, a set of fourth-order IFCs \(\{\Phi_{ijkl}\}\), where \(i, j, k\), and \(l\) label atoms in the supercell, can be categorized into four different subsets; on-site, two-body, three-body, and four-body terms. Neglecting the Cartesian coordinates of IFCs for simplicity, each subset contains the IFC elements shown as follows:

on-site

\(\{\Phi_{iiii}\}\)

two-body

\(\{\Phi_{iijj}\}\), \(\{\Phi_{iiij}\}\) (\(i\neq j\))

three-body

\(\{\Phi_{iijk}\}\) (\(i\neq j, i\neq k, j \neq k\))

four-body

\(\{\Phi_{ijkl}\}\) (all subscripts are different from each other)

Since the four-body clusters are expected to be less important than the three-body and less-body clusters, you may want to exclude the four-body terms from the Taylor expansion potential because the number of such terms is huge. This can be done by setting the NBODY tag as NBODY = 2 3 3 together with NORDER = 3.

More examples:

NORDER = 2; NBODY = 2 2 includes harmonic and cubic IFCs but excludes three-body clusters from the cubic terms.

NORDER = 5; NBODY = 2 3 3 2 2 includes anharmonic terms up to the sixth-order, where the four-body clusters are excluded from the fourth-order IFCs, and the multi (\(\geq 3\))-body clusters are excluded from the fifth- and sixth-order IFCs.


“&cutoff”-field#

In this entry field, one needs to specify cutoff radii of interaction for each order in units of bohr (or of LENGTH_UNIT if given). In the current implementation, cutoff radii should be defined for every possible pair of atomic elements. For example, the cutoff entry for a harmonic calculation (NORDER = 1) of Si should be like

&cutoff
 Si-Si 10.0
/

This means that the cutoff radius of 10 \(a_{0}\) is used for harmonic Si-Si terms. Please note that the first column should be two character strings, which are contained in the KD-tag, connected by a hyphen (’-’).

When one wants to consider cubic terms (NORDER = 2), please specify the cutoff radius for cubic terms in the third column as the following:

&cutoff
 Si-Si 10.0 5.6 # Pair r_{2} r_{3}
/

Instead of giving specific cutoff radii, one can write “None” as follows:

&cutoff
 Si-Si None 5.6
/

which means that all possible harmonic terms between Si-Si atoms will be included.

Caution

Setting ‘None’ for anharmonic terms can greatly increase the number of parameters and thereby increase the computational cost.

When there are more than two atomic elements, please specify the cutoff radii between every possible pair of atomic elements. In the case of MgO, the cutoff entry should be like

&cutoff
 Mg-Mg 8.0
 O-O 8.0
 Mg-O 10.0
/

which can equivalently be written by using the wild card (’*’) as

&cutoff
 *-* 8.0
 Mg-O 10.0 # Overwrite the cutoff radius for Mg-O harmonic interactions
/

Important

Cutoff radii specified by an earlier entry are overwritten by a new entry that comes later.

Once the cutoff radii are properly given, harmonic force constants \(\Phi_{i,j}^{\mu,\nu}\) satisfying \(r_{ij} \le r_{c}^{\mathrm{KD}[i]-\mathrm{KD}[j]}\) will be searched.

In the case of cubic terms, force constants \(\Phi_{ijk}^{\mu\nu\lambda}\) satisfying \(r_{ij} \le r_{c}^{\mathrm{KD}[i]-\mathrm{KD}[j]}\), \(r_{ik} \le r_{c}^{\mathrm{KD}[i]-\mathrm{KD}[k]}\), and \(r_{jk} \le r_{c}^{\mathrm{KD}[j]-\mathrm{KD}[k]}\) will be searched and determined by fitting.


“&cell”-field#

Please give the cell parameters in this entry in units of bohr (or of LENGTH_UNIT if given) as the following:

&cell
 a
 a11 a12 a13
 a21 a22 a23
 a31 a32 a33
/

The cell parameters are then given by \(\vec{a}_{1} = a \times (a_{11}, a_{12}, a_{13})\), \(\vec{a}_{2} = a \times (a_{21}, a_{22}, a_{23})\), and \(\vec{a}_{3} = a \times (a_{31}, a_{32}, a_{33})\).


“&position”-field#

In this field, one needs to specify the atomic element and fractional coordinate of atoms in the supercell. Each line should be

ikd xf[1] xf[2] xf[3]

where ikd is an integer specifying the atomic element in the order of the KD tag and xf[i] is the fractional coordinate of an atom. There should be NAT such lines in the &position entry field.


“&optimize”-field#

This field is necessary when MODE = optimize.

  • LMODEL-tag : Choice of the linear model used for estimating force constants

“least-squares”, “LS”, “OLS”, 1

Ordinary least square

“elastic-net”, “enet”, 2

Elastic net

“adaptive-lasso”, 3

Adaptive LASSO

Default:

least-squares

Type:

String

Description:

When LMODEL = ols, the force constants are estimated from the displacement-force datasets via the ordinary least-squares (OLS), which is usually sufficient to calculate harmonic and third-order force constants.

The elastic net (LMODEL = enet) or adaptive LASSO (LMODEL = adaptive-lasso) are useful for calculating fourth-order (and higher-order) force constants. When the elastic net or adaptive LASSO is selected, the users have to set the following related tags: CV, L1_RATIO, L1_ALPHA, CV_MAXALPHA, CV_MINALPHA, CV_NALPHA, STANDARDIZE, ENET_DNORM, MAXITER, CONV_TOL, NWRITE, SOLUTION_PATH, DEBIAS_OLS, STOP_CRITERION.


  • DFSET-tag: File name containing displacement-force datasets for training

Default:

None

Type:

String

Description:

The format of DFSET can be found here


  • NDATA-tag : Number of displacement-force data sets

Default:

None

Type:

Integer

Description:

If NDATA is not given, the code reads all lines of DFSET (excluding comment lines) and estimates NDATA by dividing the line number by NAT. If the number of lines is not divisible by NAT, an error is raised. DFSET should contain at least NDATA\(\times\) NAT lines.


  • NSTART, NEND-tags : Specifies the range of data to be used for training

Default:

NSTART = 1, NEND = NDATA

Type:

Integer

Example:

To use the data in the range of [20:30] out of 50 entries, the tags should be NSTART = 20 and NEND = 30.


  • SKIP-tag : Specifies the range of data to be skipped for training

Default:

None

Type:

Two integers connected by a hyphen

Description:

SKIP =\(i\)-\(j\) skips the data in the range of [\(i\):\(j\)]. The \(i\) and \(j\) must satisfy \(1\leq i \leq j \leq\) NDATA. This option may be useful when doing cross-validation manually (CV=-1).


  • ICONST-tag = 0 | 1 | 2 | 3 | 4 | 11 | 12 | 13 | 14

0

No constraints

1

Constraint for translational invariance is imposed between IFCs.
Available only when LMODEL = ols.

11

Same as ICONST = 1 but the constraint is imposed algebraically rather than numerically.
Select this option when LMODEL = enet or adaptive-lasso.

2

In addition to ICONST = 1, constraints for rotational invariance will be
imposed up to (NORDER + 1)th order. Available only when LMODEL = ols.

3

In addition to ICONST = 2, constraints for rotational invariance between (NORDER + 1)th order
and (NORDER + 2)th order, which are zero, will be considered.
Available only when LMODEL = ols.

4

In addition to ICONST = 2, constraints for the Huang invariance will be imposed.
Available only when LMODEL = ols. For a stressed reference structure, see STRESS.

12

Same as ICONST = 2 but imposed algebraically. Usable with any LMODEL.

13

Same as ICONST = 3 but imposed algebraically. Usable with any LMODEL.

14

Same as ICONST = 4 but imposed algebraically. Usable with any LMODEL.

Note

When LMODEL = enet or adaptive-lasso, ICONST = 1, 2, 3, 4 are not rejected but are treated as ICONST = 11, 12, 13, 14, respectively, i.e., the constraints are imposed algebraically.

Note

With ICONST = 12, 13, 14, the inter-order rotational constraints make some higher-order IFCs linear functions of lower-order ones. These dependent IFCs are eliminated before the fit, so they are not penalized by the L1 term of LMODEL = enet or adaptive-lasso. As in ICONST = 2, 3, 4, inter-order constraints involving an order fixed by FC2FIX or FC3FIX are not imposed.

Note

The rotational constraints (ICONST = 2, 3, 4, 12, 13, 14) are built from the Cartesian atomic positions, so the input structure must be symmetric to nearly machine precision. If the lattice vectors or positions are given with only about 10 digits, the inter-order constraints have no clear rank, and the number of free parameters and the resulting IFCs may depend on ICONST, ALGO_REDUCTION, and TOL_CONST. Give the structure with full double precision or set SYMMETRIZE_STRUCTURE = 1.

Even with an exact structure, a structure with internal relaxation (atoms off high-symmetry positions) combined with a short cutoff of the highest order can give nearly dependent rotational constraints. ICONST = 2, 3, 4 and, with the default ALGO_REDUCTION = 3, ICONST = 12, 13, 14 drop those weaker than TOL_CONST (relative, row by row). The other ALGO_REDUCTION options, and ALGO_REDUCTION = 3 for constraint blocks too large for the dense QR in ALM built without USE_SUITESPARSE_BACKEND, decide the rank in several Gauss-Jordan stages and may keep some of them, which shows up as a fitting error larger than with ICONST = 2, 3, 4. In that case, use ICONST = 2, 3, 4 (with LMODEL = ols), or increase TOL_CONST or the cutoff radius.

Default:

11

Type:

Integer

Description:

See this page for the numerical formulae.


  • PERIODIC_IMAGE_CONV-tag = 0 | 1

0

Impose the constraints on IFCs (acoustic sum rule) in the considering supercell.

1

Consider the periodic images when generating the constraints.
The resultant IFCs simultaneously satisfy the permutation symmetry, ASR,
and the space group symmetry in the infinite space.
For more details, please see Appendix D of the original paper.
(Note that we use the term “mirror image” instead of “periodic image” in the paper.)
Default:

1

Type:

Integer


  • ROTAXIS-tag : Rotation axis used to estimate constraints for rotational invariance. This entry is necessary when ICONST = 2, 3, 4, 12, 13, 14.

Default:

None

Type:

String

Example:

When one wants to consider the rotational invariance around the \(x\)-axis, one should give ROTAXIS = x. If one needs additional constraints for the rotation around the \(y\)-axis, ROTAXIS should be ROTAXIS = xy.


  • STRESS-tag : Cauchy stress \(\sigma\) of the reference structure (in GPa) used to generalize the Huang invariance

Default:

None (stress-free Huang invariance)

Type:

Array of double (6 or 9 values)

Description:

The Huang invariance imposed with ICONST = 4 or 14 assumes a stress-free reference structure. When the reference structure is under a stress \(\sigma\), the long-wavelength condition becomes

\[[\alpha\beta,\gamma\delta] - [\gamma\delta,\alpha\beta] = \sigma_{\gamma\delta}\delta_{\alpha\beta} - \sigma_{\alpha\beta}\delta_{\gamma\delta}, \quad [\alpha\beta,\gamma\delta] = -\frac{1}{2\Omega}\sum_{\kappa,\ell\kappa'} \Phi_{\alpha\beta}(0\kappa;\ell\kappa')\, x_{\gamma} x_{\delta},\]

where \(\boldsymbol{x} = \boldsymbol{r}(\ell\kappa') - \boldsymbol{r}(0\kappa)\) and \(\Omega\) is the volume of the primitive cell. STRESS gives \(\sigma\) either as 9 values (the \(3\times 3\) tensor, row by row) or as 6 Voigt values in the order xx yy zz yz xz xy, in the Cartesian frame of the input lattice vectors.

The sign convention is tension positive. Note that the stress printed by VASP (in kB line of OUTCAR) is positive under compression: convert it as \(\sigma = -0.1 \times\) (VASP value) GPa. ASE’s get_stress() is already tension positive (in eV/Å3). LAMMPS reports the pressure tensor, so \(\sigma = -p\).

The input tensor is symmetrized, \((\sigma + \sigma^{T})/2\), and averaged over the point group of the structure; the input and symmetrized tensors are printed, and a warning is shown when the removed part exceeds 1% of \(|\sigma|\). A hydrostatic stress does not change the Huang conditions. STRESS is ignored (with a warning) unless ICONST = 4 or 14.

Example:

STRESS = 1.2 0.1 0.7 0 0 0.2 (Voigt, GPa)


  • ALGO_REDUCTION-tag = 0 | 1 | 2 | 3 | 4

0

No separate reduction step; the elimination map is built by sparse
Gauss-Jordan elimination, i.e., the same result as ALGO_REDUCTION = 1

1

Sparse Gauss-Jordan elimination (legacy)

2

Rank-revealing QR row selection followed by sparse Gauss-Jordan elimination

3

Rank-revealing QR row selection with the scale-free TOL_CONST test, followed by sparse
Gauss-Jordan elimination with threshold pivoting. Higher-order IFCs are expressed through
lower-order ones where the pivot size allows; the log reports how many relations go the other
way. Without USE_SUITESPARSE_BACKEND, blocks above 5e7 matrix elements use the Gauss-Jordan
elimination with partial pivoting of earlier versions.

4

Reduced echelon form from a sparse Householder QR (SuiteSparseQR) with scale-aware rank
detection. Requires ALM built with USE_SUITESPARSE_BACKEND.
Default:

3

Type:

Integer

Description:

Algorithm used to reduce the constraints to an elimination map when ICONST >= 11. All options keep the independent parameters as individual IFCs; ICONST = 1, 2, 3, 4 always use a rank-revealing QR.


  • FC2FIX-tag : File to which the harmonic terms are fixed upon training

Default:

None

Type:

String

Description:

When the FC2FIX-tag is given, harmonic force constants are fixed to the values stored in the FC2FIX file. This may be useful for optimizing cubic and higher-order terms without changing the harmonic terms. Please make sure that the number of harmonic terms in the new computational condition is the same as that in the FC2FIX file.

Note

FC2XML is a deprecated alias of FC2FIX kept for backward compatibility. It will be removed in a future release, so please use FC2FIX instead.

Important

When FC2FIX is an XML file, the FCSYM_BASIS option must be the same as the one used when creating it. If FCSYM_BASIS is not given, the basis recorded in the XML file is used, so files created with the old default (Lattice, before version 2.0) still work. An explicitly given FCSYM_BASIS that differs from the file raises an error. An HDF5 (.h5) file can be used with either basis.


  • FC3FIX-tag : File to which the cubic terms are fixed upon training

Default:

None

Type:

String

Description:

Same as the FC2FIX-tag, but FC3FIX is used to fix cubic force constants. (FC3XML is a deprecated alias of FC3FIX.)

Important

When FC3FIX is an XML file, the FCSYM_BASIS option must be the same as the one used when creating it. If FCSYM_BASIS is not given, the basis recorded in the XML file is used, so files created with the old default (Lattice, before version 2.0) still work. An explicitly given FCSYM_BASIS that differs from the file raises an error. An HDF5 (.h5) file can be used with either basis.


  • SPARSE-tag = 0 | 1

0

Use a direct solver (SVD or QRD) to estimate force constants

1

Use a sparse solver to estimate force constants

Default:

0

Type:

Integer

Description:

When you want to calculate force constants of a large system and generate training datasets by displacing only a few atoms from equilibrium positions, the resulting sensing matrix becomes large but sparse. For such matrices, a sparse solver is expected to be more efficient than SVD or QRD in terms of both memory usage and computational time. Effective when LMODEL = ols.

When to use: prefer SPARSE = 1 for large systems whose sensing matrix is genuinely sparse (only a handful of atoms displaced per configuration). For small or dense problems the default direct solvers (SPARSE = 0) are usually faster, and the sparse factorization can even use more memory because of fill-in. When the sensing matrix is dense but large, consider USE_CHOLESKY instead.

Note

The sparse path has two different solver routes. With algebraic constraints (ICONST >= 10; the default is ICONST = 11), the reduced sparse least-squares problem is solved by the SPARSESOLVER selected below. With constraints imposed numerically (ICONST = 1, 2, 3, 4), the sparse path instead assembles the symmetric-indefinite KKT system. The merged constraint matrix is first reduced to full row rank by a rank-revealing QR (essential: a rank-deficient constraint matrix makes the KKT singular and the fitted IFCs violate the acoustic-sum-rule / rotational invariance). The KKT solver is then chosen by build: Intel MKL PARDISO (-DUSE_MKL_BACKEND=yes) or Apple Accelerate (-DUSE_ACCEL_BACKEND=yes) use a symmetric-indefinite LDLT; SuiteSparse (-DUSE_SUITESPARSE_BACKEND=yes) adds the multithreaded SuiteSparseQR fallback; otherwise Eigen’s serial SparseLU and SparseQR fallbacks are used. See Optional linear-algebra backends on the installation page.


  • SPARSESOLVER-tag : Type of the sparse solver to use

Default:

SimplicialLDLT

Type:

String

Description:

Selects the sparse solver used for the unconstrained or algebraically constrained sparse OLS fit (LMODEL = ols, SPARSE = 1, and ICONST >= 10). Available options are SimplicialLDLT (sparse Cholesky on the normal matrix, direct), SparseQR (direct sparse QR), and the iterative Krylov solvers ConjugateGradient, LeastSquaresConjugateGradient, and BiCGSTAB. When an iterative solver is selected, the stopping criterion is controlled by the CONV_TOL and MAXITER tags. As a rule of thumb, use SimplicialLDLT (the default) for well-conditioned problems, SparseQR when the direct factorization fails or the matrix is poorly conditioned, and the iterative solvers when even the sparse factorization is too large to store.

Two additional high-performance solvers are available when ALM is built with the SuiteSparse backend (-DUSE_SUITESPARSE_BACKEND=yes; see Optional linear-algebra backends): SuiteSparseQR, a multithreaded rank-revealing multifrontal sparse QR that factorizes the rectangular sensing matrix directly (a faster, parallel alternative to SparseQR for large or ill-conditioned problems), and CHOLMOD, a supernodal sparse Cholesky on the normal matrix \(A^{\mathsf T}A\) (a faster alternative to SimplicialLDLT for well-conditioned problems). Selecting either of these without building the SuiteSparse backend stops with an error.

With numerically imposed constraints (ICONST = 1, 2, 3, 4), most values of SPARSESOLVER do not select the KKT direct solver. The code first tries the KKT backend compiled into the executable (MKL/PARDISO or Apple/Accelerate LDLT), then SuiteSparseQR if the SuiteSparse backend is enabled, then Eigen’s built-in sparse fallbacks. MINRES is the special exception (experimental): selecting SPARSESOLVER = MINRES chooses an iterative solver for this KKT path. It never forms the dense factor used by direct solvers, so it can be the memory-feasible choice when direct KKT factorizations exhaust memory because the constraint matrix has dense rows (the rotational-invariance constraints of ICONST = 2, 3). It is preconditioned by a block-diagonal SPD preconditioner (Jacobi on \(A^{\mathsf T}A\) plus the dense constraint Schur complement); CONV_TOL and MAXITER control the stopping criterion. After the solve it checks the true KKT residual and \|C x - d\| and aborts if they are not met, rather than returning a solution that would violate the sum rules. On badly-scaled or extremely ill-conditioned KKT systems the simple block-diagonal preconditioner may be insufficient and MINRES will report non-convergence – in that case raise MAXITER / loosen CONV_TOL, or use a direct LDLT backend on a reduced problem.

SPARSESOLVER is ignored by USE_CHOLESKY = 1. In that mode the code forms \(A^{\mathsf T}A\) directly and uses a sparse Cholesky factorization (CHOLMOD when the SuiteSparse backend is available, otherwise SimplicialLDLT).

See also

Eigen documentation page: Solving Sparse Linear Systems


  • USE_CHOLESKY-tag = 0 | 1

0

Use a direct solver (SVD or QRD) to estimate force constants

1

Solve the normal equations by Cholesky decomposition

Default:

0

Type:

Integer

Description:

When USE_CHOLESKY = 1, the ordinary least-squares fit is obtained by forming the normal equations \(A^{\mathsf{T}} A\,\boldsymbol{\Phi} = A^{\mathsf{T}} \boldsymbol{b}\) and factorizing \(A^{\mathsf{T}} A\) with a Cholesky decomposition (a dense LLT, or a sparse Cholesky factorization when SPARSE = 1: CHOLMOD with the SuiteSparse backend, otherwise SimplicialLDLT). This is typically about twice as fast and uses noticeably less memory than the default SVD/QRD direct solvers, which makes it attractive for large, well-conditioned overdetermined fits (many more rows than columns). The trade-off is accuracy: forming \(A^{\mathsf{T}} A\) squares the condition number of \(A\), so the solution degrades when the sensing matrix is ill-conditioned. Keep the default USE_CHOLESKY = 0 (SVD/QRD) when robustness matters or the fit is close to rank-deficient. Effective when LMODEL = ols.

Important

USE_CHOLESKY = 1 cannot be combined with the numerically imposed constraints ICONST = 1, 2, 3, 4. When constraints are required, impose them algebraically with ICONST = 11, 12, 13, 14 instead; otherwise the code stops with an error.


  • CHUNKSIZE-tag : Number of configurations processed per pass when building \(A^{\mathsf{T}} A\)

Default:

100

Type:

Integer

Description:

Relevant only when USE_CHOLESKY = 1. Instead of forming the full dense sensing matrix \(A\) (size \(N_{\mathrm{rows}} \times N_{\mathrm{cols}}\)) at once, the normal-equations matrix \(A^{\mathsf{T}} A\) is accumulated in chunks of CHUNKSIZE training configurations: for each chunk only a small block of \(A\) is built and added into \(A^{\mathsf{T}} A\). This caps the peak memory at roughly the larger of \(A^{\mathsf{T}} A\) itself and one chunk’s block (proportional to CHUNKSIZE \(\times\, N_{\mathrm{cols}}\)), rather than the full \(A\). A smaller CHUNKSIZE lowers the peak memory but uses more passes with smaller, less efficient matrix products; a larger CHUNKSIZE is faster but needs more memory. Lowering it below the point where one chunk is already smaller than \(A^{\mathsf{T}} A\) gives no further savings, since \(A^{\mathsf{T}} A\) sets the memory floor. The default (100) is a reasonable balance; reduce it if the chunk block does not fit in memory.


  • NWRITE-tag : Interval of the progress output of iterative L1 solvers

Default:

1000

Type:

Integer

Description:

The progress of the L1 solver iteration (LMODEL = enet | adaptive-lasso) is printed every NWRITE iterations when VERBOSITY >= 2 is set in the &general field.


  • MAXITER-tag : Number of maximum iterations in iterative algorithms

Default:

10,000

Type:

Integer

Description:

Effective when an iterative solver is selected via SPARSESOLVER (when LMODEL = ols) or when LMODEL = enet | adaptive-lasso.


  • CONV_TOL-tag : Convergence criterion of iterative algorithms

Default:

1.0e-8

Type:

Double

Description:

When LMODEL = ols and an iterative solver is selected via SPARSESOLVER, CONV_TOL value is passed to the Eigen3 function via setTolerance(). When LMODEL = enet | adaptive-lasso, the L1 solver iteration stops at \(i\)th iteration if \(\sqrt{\frac{1}{N}|\boldsymbol{\Phi}_{i} - \boldsymbol{\Phi}_{i-1}|_{2}^{2}} <\) CONV_TOL is satisfied, where \(N\) is the length of the vector \(\boldsymbol{\Phi}\).

See also

Eigen documentation page: IterativeSolverBase


  • L1_RATIO-tag : The ratio of the L1 regularization term

Default:

1.0 (LASSO)

Type:

Double

Description:

The L1_RATIO changes the regularization term as L1_ALPHA \(\times\) [L1_RATIO \(|\boldsymbol{\Phi}|_{1}\) + \(\frac{1}{2}\) (1-L1_RATIO) \(|\boldsymbol{\Phi}|_{2}^{2}\)]. Therefore, L1_RATIO = 1 corresponds to LASSO. L1_RATIO must be 0 < L1_ratio <= 1. Effective when LMODEL = enet. See also here.


  • L1_SOLVER-tag = cd | fista | admm

cd

Use the coordinate descent solver.

fista

Use the fast iterative shrinkage-thresholding algorithm.

admm

Use the alternating direction method of multipliers (experimental).

Default:

cd

Type:

String

Description:

Selects the iterative solver used when LMODEL = enet | adaptive-lasso. All three solvers target the same elastic-net or adaptive-LASSO objective and are interchangeable. The cd solver is the coordinate descent method; it warm-starts along the cross-validation alpha path and is the fastest choice for that path, so it is the recommended default. The fista solver avoids forming the dense \(N \times N\) Gram matrix and can reduce memory usage for large dense L1-regularized problems. The admm solver is provided mainly for research and cross-checking; on warm-started alpha paths it is generally slower than cd and offers no advantage for the typical problem sizes here.


  • L1_ALPHA-tag : The coefficient of the L1 regularization term

Default:

0.0

Type:

Double

Description:

This tag is used when LMODEL = enet | adaptive-lasso and CV = 0. See also here.


  • CV-tag : Cross-validation mode for elastic net

0

Cross-validation mode is off.
The elastic net optimization is solved with the given L1_ALPHA value.
The force constants are written to PREFIX.h5 (and PREFIX.xml / PREFIX.fcs with FCS_ALAMODE = 1).

>= 2

CV-fold cross-validation is performed automatically.
NDATA training datasets are divided into CV subsets, and CV different combinations of
training-validation datasets are created internally. For each combination, the elastic net
optimization is solved with the various L1_ALPHA values defined by the CV_MINALPHA,
CV_MAXALPHA, and CV_NALPHA tags. The result of each cross-validation is stored in
PREFIX.cvset[1, …, CV], and their average and deviation are stored in PREFIX.cvscore.

-1

The cross-validation is performed manually.
The Taylor expansion potential is trained by using the training datasets in DFSET, and
the validation score is calculated by using the data in DFSET_CV for various L1_ALPHA values
defined by the CV_MINALPHA, CV_MAXALPHA, and CV_NALPHA tags.
After the calculation, the fitting and validation errors are stored in PREFIX.cvset.
This option may be convenient for a large-scale problem since multiple optimization tasks with
different training-validation datasets can be done in parallel.
Default:

0

Type:

Integer

Description:

This tag is used when LMODEL = enet | adaptive-lasso.


  • DFSET_CV-tag : File name containing displacement-force datasets used for manual cross-validation

Default:

DFSET_CV = DFSET

Type:

String

Description:

This tag is used when LMODEL = enet | adaptive-lasso and CV = -1.


  • NDATA_CV-tag : Number of displacement-force validation datasets

Default:

None

Type:

Integer

Description:

This tag is used when LMODEL = enet | adaptive-lasso and CV = -1.


  • NSTART_CV, NEND_CV-tags : Specifies the range of data to be used for validation

Default:

NSTART_CV = 1, NEND_CV = NDATA_CV

Type:

Integer

Description:

This tag is used when LMODEL = enet | adaptive-lasso and CV = -1.


  • CV_MINALPHA, CV_MAXALPHA, CV_NALPHA-tags : Options to specify the L1_ALPHA values used in cross-validation

Default:

CV_MAXALPHA is set automatically

CV_MINALPHA = CV_MAXALPHA * 1.0e-6

CV_NALPHA = 50

Type:

Double, Double, Integer

Description:

CV_NALPHA values of L1_ALPHA are generated from CV_MINALPHA to CV_MAXALPHA in logarithmic scale. When CV_MAXALPHA is not specified by user, the code automatically sets CV_MAXALPHA so that the maximum L1_ALPHA makes all coefficients zero. The default value of CV_MINALPHA is CV_MAXALPHA * 1.0e-6, which is reasonable in many cases. If the minimum value of the validation score is found at CV_MINALPHA, you may need to use a smaller value of CV_MINALPHA. This tag is used when LMODEL = enet | adaptive-lasso and the cross-validation mode is on (CV > 0 or CV = -1).


  • STANDARDIZE-tag = 0 | 1

0

Do not standardize the sensing matrix

1

Each column of the sensing matrix is standardized in such a way that its mean value
becomes 0 and standard deviation becomes 1.
Default:

1

Type:

Integer

Description:

This option influences the optimal L1_ALPHA value. So, if you change the STANDARDIZE option, you have to rerun the cross-validation. Effective when LMODEL = enet | adaptive-lasso.


  • ENET_DNORM-tag : Normalization factor of atomic displacements

Default:

1.0

Type:

Double

Description:

The normalization factor of atomic displacement \(u_{0}\) in units of bohr. When \(u_{0} (\neq 1)\) is given, the displacement data are scaled as \(u_{i} \rightarrow u_{i}/u_{0}\) before constructing the sensing matrix. This option influences the optimal L1_ALPHA value. So, if you change the ENET_DNORM value, you will have to rerun the cross-validation. Effective only when LMODEL = enet and STANDARDIZE = 0.


  • SOLUTION_PATH-tag = 0 | 1

0

Do not save the solution path.

1

Save the solution path in PREFIX.solution_path (CV = -1) or, for each
cross-validation subset, in PREFIX.solution_path1, …, PREFIX.solution_pathN (CV >= 2).
Default:

0

Type:

Integer

Description:

Effective when LMODEL = enet | adaptive-lasso and the cross-validation mode is on.


  • DEBIAS_OLS-tag = 0 | 1

0

Save the solution of the L1-regularized problem to the force-constant files.

1

After the solution of the elastic net or adaptive LASSO optimization problem is obtained,
only non-zero coefficients are collected, and the ordinary least-squares fitting is
solved again with the non-zero coefficients before saving the results to the force-constant
files. This might be useful to reduce the bias of the L1-regularized solution.
Default:

0

Type:

Integer

Description:

Effective when LMODEL = enet | adaptive-lasso and CV = 0.


  • STOP_CRITERION-tag : The scan over L1_ALPHA stops when the cross-validation score keeps increasing in STOP_CRITERION consecutive steps

Default:

5

Type:

Integer

Description:

Effective when LMODEL = enet | adaptive-lasso and the cross-validation mode is turned on (CV > 0 or CV = -1).


How to make a DFSET file#

Format of DFSET#

The displacement-force data sets obtained by first-principles (or classical force-field) calculations have to be saved to a file, say DFSET. Then, the force constants are estimated by setting DFSET = DFSET and with MODE = optimize.

The DFSET file must contain the atomic displacements and corresponding forces in Cartesian coordinate for at least NDATA structures (displacement patterns) in the following format (one line per atom, blocks separated by an optional comment line beginning with #):

# Structure number 1 (this is just a comment line)
u_x(1)     u_y(1)     u_z(1)     f_x(1)     f_y(1)     f_z(1)
u_x(2)     u_y(2)     u_z(2)     f_x(2)     f_y(2)     f_z(2)
 ...        ...        ...        ...        ...        ...
u_x(NAT)   u_y(NAT)   u_z(NAT)   f_x(NAT)   f_y(NAT)   f_z(NAT)
# Structure number 2
u_x(1)     u_y(1)     u_z(1)     f_x(1)     f_y(1)     f_z(1)
 ...        ...        ...        ...        ...        ...

Here, NAT is the number of atoms in the supercell. The unit of displacements and forces must be bohr and Ry/bohr, respectively, unless other units are declared with the LENGTH_UNIT and FORCE_UNIT tags (e.g., LENGTH_UNIT = angstrom; FORCE_UNIT = eV/angstrom for data taken directly from VASP output).