Skip to content

INCAR key reference

Every key mliport accepts in an INCAR.mliport file, what it means, its default and the values it accepts. The tables are generated from the option table in mliport/config/schema.py - the same table the templates and mliport config schema come from - so this page cannot drift from the code.

Three things apply everywhere on this page. A key is written as KEY = VALUE, one assignment per line, # starts a comment, and the file name is INCAR.mliport (mliport never reads a file called plain INCAR, which may be a real VASP input). REQUIRED is a placeholder: the template cannot invent a model checkpoint or a start structure, so you fill it in. Precedence is CLI, then a profile, then a model alias, then the INCAR, then settings.ini, then the built-in default; see Configuration for the layering and the task-folder usage.

Units are the ones named in each row: femtoseconds (fs), picoseconds (ps), Angstrom, electronvolt (eV), Kelvin and gigapascal (GPa). Energies are MLIP total energies and are only comparable within one model identity.

Keys common to every calculation type

Calculation

Key Type Default Allowed values Meaning Applies when Aliases
CALC_TYPE str - sp, opt, md, neb, batch, confsearch Top-level calculation type (selects the runner). - CALCULATION

Model

Key Type Default Allowed values Meaning Applies when Aliases
MODEL_TYPE str - uma, fairchem, mace, dpa, grace, nep MLIP engine. - -
MODEL_PATH str required (REQUIRED) - Model checkpoint file or directory. A value without any path separator is a file name: it is resolved next to the declaring file first, then in /models (one level of subdirectories allowed). - -
TASK str - - UMA task (omat/omol/...) or bulk/molecule for other engines. - -
INFERENCE_MODE str - default, turbo UMA inference mode. Non-default modes are rejected for other engines. - -
DEFAULT_DTYPE str - float32, float64 MACE model dtype (accuracy-first default: float64). - -
HEAD str - - MACE or DeepMD model head/branch (multi-task models). - -
ACTIVATION_CHECKPOINTING bool - - GPU memory saving (UMA, overrides inference_mode preset). - -
GPU_MEMORY_GROWTH bool - - GRACE TensorFlow allocator grows on demand instead of claiming the whole visible GPU. - -
GPU_MEMORY_LIMIT_MB int - 1.. Hard GRACE TensorFlow logical-device memory limit in MiB. - -
NEIGHBOR_CACHE bool - - GRACE verlet-style neighbor-list cache (extended cutoff + exact re-filter). The complete periodic-image multiset matches a fresh search; pair ordering may differ. Default: True. - -
NEIGHBOR_SKIN float - 1e-12.. GRACE neighbor-cache skin in Å: the cached neighbor list is rebuilt when any atom moves more than skin/2. Larger values rebuild less often but filter longer lists. - -
D3 bool - - Add Grimme DFT-D3 dispersion (torch-dftd) on top of the model. Only for models trained on dispersion-free reference data. - -
D3_XC str - - Reference functional of the model's training data, selecting the D3 parameters (default: pbe). - -
D3_DAMPING str - zero, bj, zerom, bjm DFT-D3 damping function (default: bj, i.e. D3(BJ)). - -
DPA_RUNTIME str - current, legacy DPA runtime environment: 'current' (pinned DPA-4-capable deepmd-kit) or 'legacy' (.venv-dpa-legacy, deepmd-kit 3.0.0b3, for DPA-2 checkpoints such as DPA-SSE). No automatic fallback. - -
NEP_RUNTIME str - ase, gpumd NEP execution runtime: 'ase' (Calorine CPUNEP/GPUNEP calculator) or 'gpumd' (native gpumd process). Required for MODEL_TYPE=nep; mliport never infers it or falls back between the two. - -
NEP_GPUMD_COMMAND str - - Path to the gpumd executable for NEP_RUNTIME=gpumd (and for the GPUNEP ASE adapter). Launched as an argv list, never through a shell; its real path, version and SHA256 are recorded. - -

Device and resources

Key Type Default Allowed values Meaning Applies when Aliases
DEVICE str cpu - Compute device: cpu, cuda, gpu or cuda:N. - -
TORCH_NUM_THREADS int - 1.. Backend CPU intra-op threads (PyTorch for UMA/MACE/DPA .pt; TensorFlow for GRACE; legacy canonical field name). - CPU_THREADS

Molecular dynamics

Key Type Default Allowed values Meaning Applies when Aliases
FMAX_ABORT float - 0.0.. MD force-safety abort threshold in eV/Angstrom. - -
SEED int - 0.. Random seed for reproducible MD. - DEFAULT_SEED

Output

Key Type Default Allowed values Meaning Applies when Aliases
WRITE_RESOLVED_CONFIG bool .TRUE. - Write resolved_config.json before a calculation. - -
WRITE_FORCES bool .TRUE. - Write forces to OUTCAR. - -
WRITE_OUTCAR bool - - Write the VASP-like OUTCAR interoperability view. - -
WRITE_XDATCAR bool - - Write the VASP XDATCAR interoperability trajectory. - -
WRITE_STRESS bool .TRUE. - Write stress to OUTCAR. - -
WRITE_TRAJECTORY bool - - Write ASE trajectory. - -
WRITE_JSON bool - - Write JSON results. - -
OUTPUT_FORMAT str VASP - Output format (VASP). - -
OUTPUT_DIR str - - Output directory for the calculation. Used when the command line does not pass --output; the resolved config is written into the final job directory. Relative values resolve against the file that declares them (the INCAR folder for an INCAR), and mliport run defaults to /results when neither --output nor OUTPUT_DIR is given. - -
JOB_NAME str - - Optional job name. - -

General

Key Type Default Allowed values Meaning Applies when Aliases
STRICT_CONFIG bool .TRUE. - Treat unknown configuration keys as errors. - -

Single point

CALC_TYPE = SP evaluates energy, forces and stress of one structure. Output control keys apply here as they do for optimization and MD.

Calculation

Key Type Default Allowed values Meaning Applies when Aliases
CALC_TYPE str - sp, opt, md, neb, batch, confsearch Top-level calculation type (selects the runner). - CALCULATION

Model

Key Type Default Allowed values Meaning Applies when Aliases
MODEL_TYPE str - uma, fairchem, mace, dpa, grace, nep MLIP engine. - -
MODEL_PATH str required (REQUIRED) - Model checkpoint file or directory. A value without any path separator is a file name: it is resolved next to the declaring file first, then in /models (one level of subdirectories allowed). - -
TASK str - - UMA task (omat/omol/...) or bulk/molecule for other engines. - -
INFERENCE_MODE str default default, turbo UMA inference mode. Non-default modes are rejected for other engines. - -
DEFAULT_DTYPE str - float32, float64 MACE model dtype (accuracy-first default: float64). - -
HEAD str - - MACE or DeepMD model head/branch (multi-task models). - -
ACTIVATION_CHECKPOINTING bool - - GPU memory saving (UMA, overrides inference_mode preset). - -
GPU_MEMORY_GROWTH bool - - GRACE TensorFlow allocator grows on demand instead of claiming the whole visible GPU. - -
GPU_MEMORY_LIMIT_MB int - 1.. Hard GRACE TensorFlow logical-device memory limit in MiB. - -
NEIGHBOR_CACHE bool - - GRACE verlet-style neighbor-list cache (extended cutoff + exact re-filter). The complete periodic-image multiset matches a fresh search; pair ordering may differ. Default: True. - -
NEIGHBOR_SKIN float - 1e-12.. GRACE neighbor-cache skin in Å: the cached neighbor list is rebuilt when any atom moves more than skin/2. Larger values rebuild less often but filter longer lists. - -
D3 bool - - Add Grimme DFT-D3 dispersion (torch-dftd) on top of the model. Only for models trained on dispersion-free reference data. - -
D3_XC str - - Reference functional of the model's training data, selecting the D3 parameters (default: pbe). - -
D3_DAMPING str - zero, bj, zerom, bjm DFT-D3 damping function (default: bj, i.e. D3(BJ)). - -
DPA_RUNTIME str - current, legacy DPA runtime environment: 'current' (pinned DPA-4-capable deepmd-kit) or 'legacy' (.venv-dpa-legacy, deepmd-kit 3.0.0b3, for DPA-2 checkpoints such as DPA-SSE). No automatic fallback. - -
NEP_RUNTIME str - ase, gpumd NEP execution runtime: 'ase' (Calorine CPUNEP/GPUNEP calculator) or 'gpumd' (native gpumd process). Required for MODEL_TYPE=nep; mliport never infers it or falls back between the two. - -
NEP_GPUMD_COMMAND str - - Path to the gpumd executable for NEP_RUNTIME=gpumd (and for the GPUNEP ASE adapter). Launched as an argv list, never through a shell; its real path, version and SHA256 are recorded. - -
CHARGE int - -100..100 Total molecular charge stored in atoms.info. - -
SPIN int - 0..100 Molecular spin metadata; UMA omol uses spin multiplicity (2S+1). - SPIN_MULTIPLICITY

Device and resources

Key Type Default Allowed values Meaning Applies when Aliases
DEVICE str cpu - Compute device: cpu, cuda, gpu or cuda:N. - -
TORCH_NUM_THREADS int - 1.. Backend CPU intra-op threads (PyTorch for UMA/MACE/DPA .pt; TensorFlow for GRACE; legacy canonical field name). - CPU_THREADS

Molecular dynamics

Key Type Default Allowed values Meaning Applies when Aliases
FMAX_ABORT float - 0.0.. MD force-safety abort threshold in eV/Angstrom. - -
SEED int - 0.. Random seed for reproducible MD. - DEFAULT_SEED

Output

Key Type Default Allowed values Meaning Applies when Aliases
WRITE_RESOLVED_CONFIG bool .TRUE. - Write resolved_config.json before a calculation. - -
WRITE_FORCES bool .TRUE. - Write forces to OUTCAR. - -
WRITE_OUTCAR bool - - Write the VASP-like OUTCAR interoperability view. - -
WRITE_XDATCAR bool - - Write the VASP XDATCAR interoperability trajectory. - -
WRITE_STRESS bool .TRUE. - Write stress to OUTCAR. - -
WRITE_TRAJECTORY bool - - Write ASE trajectory. - -
WRITE_JSON bool - - Write JSON results. - -
OUTPUT_FORMAT str VASP - Output format (VASP). - -
OUTPUT_DIR str - - Output directory for the calculation. Used when the command line does not pass --output; the resolved config is written into the final job directory. Relative values resolve against the file that declares them (the INCAR folder for an INCAR), and mliport run defaults to /results when neither --output nor OUTPUT_DIR is given. - -
JOB_NAME str - - Optional job name. - -

General

Key Type Default Allowed values Meaning Applies when Aliases
STRICT_CONFIG bool .TRUE. - Treat unknown configuration keys as errors. - -

Structure optimization

CALC_TYPE = OPT relaxes the structure, with the cell optionally free.

Calculation

Key Type Default Allowed values Meaning Applies when Aliases
CALC_TYPE str - sp, opt, md, neb, batch, confsearch Top-level calculation type (selects the runner). - CALCULATION

Model

Key Type Default Allowed values Meaning Applies when Aliases
MODEL_TYPE str - uma, fairchem, mace, dpa, grace, nep MLIP engine. - -
MODEL_PATH str required (REQUIRED) - Model checkpoint file or directory. A value without any path separator is a file name: it is resolved next to the declaring file first, then in /models (one level of subdirectories allowed). - -
TASK str - - UMA task (omat/omol/...) or bulk/molecule for other engines. - -
INFERENCE_MODE str default default, turbo UMA inference mode. Non-default modes are rejected for other engines. - -
DEFAULT_DTYPE str - float32, float64 MACE model dtype (accuracy-first default: float64). - -
HEAD str - - MACE or DeepMD model head/branch (multi-task models). - -
ACTIVATION_CHECKPOINTING bool - - GPU memory saving (UMA, overrides inference_mode preset). - -
GPU_MEMORY_GROWTH bool - - GRACE TensorFlow allocator grows on demand instead of claiming the whole visible GPU. - -
GPU_MEMORY_LIMIT_MB int - 1.. Hard GRACE TensorFlow logical-device memory limit in MiB. - -
NEIGHBOR_CACHE bool - - GRACE verlet-style neighbor-list cache (extended cutoff + exact re-filter). The complete periodic-image multiset matches a fresh search; pair ordering may differ. Default: True. - -
NEIGHBOR_SKIN float - 1e-12.. GRACE neighbor-cache skin in Å: the cached neighbor list is rebuilt when any atom moves more than skin/2. Larger values rebuild less often but filter longer lists. - -
D3 bool - - Add Grimme DFT-D3 dispersion (torch-dftd) on top of the model. Only for models trained on dispersion-free reference data. - -
D3_XC str - - Reference functional of the model's training data, selecting the D3 parameters (default: pbe). - -
D3_DAMPING str - zero, bj, zerom, bjm DFT-D3 damping function (default: bj, i.e. D3(BJ)). - -
DPA_RUNTIME str - current, legacy DPA runtime environment: 'current' (pinned DPA-4-capable deepmd-kit) or 'legacy' (.venv-dpa-legacy, deepmd-kit 3.0.0b3, for DPA-2 checkpoints such as DPA-SSE). No automatic fallback. - -
NEP_RUNTIME str - ase, gpumd NEP execution runtime: 'ase' (Calorine CPUNEP/GPUNEP calculator) or 'gpumd' (native gpumd process). Required for MODEL_TYPE=nep; mliport never infers it or falls back between the two. - -
NEP_GPUMD_COMMAND str - - Path to the gpumd executable for NEP_RUNTIME=gpumd (and for the GPUNEP ASE adapter). Launched as an argv list, never through a shell; its real path, version and SHA256 are recorded. - -
CHARGE int - -100..100 Total molecular charge stored in atoms.info. - -
SPIN int - 0..100 Molecular spin metadata; UMA omol uses spin multiplicity (2S+1). - SPIN_MULTIPLICITY

Device and resources

Key Type Default Allowed values Meaning Applies when Aliases
DEVICE str cpu - Compute device: cpu, cuda, gpu or cuda:N. - -
TORCH_NUM_THREADS int - 1.. Backend CPU intra-op threads (PyTorch for UMA/MACE/DPA .pt; TensorFlow for GRACE; legacy canonical field name). - CPU_THREADS

Optimization

Key Type Default Allowed values Meaning Applies when Aliases
OPT_ALGO str FIRE FIRE, BFGS, LBFGS Geometry optimization algorithm. - -
FMAX float 0.05 0.0.. Force convergence threshold in eV/Angstrom. - -
MAX_STEPS int 500 0.. Maximum optimization steps. - -
CELL_OPT bool .FALSE. - Optimize cell parameters (requires stress support). - -
FIX_SYMMETRY bool .FALSE. - Preserve crystal symmetry during optimization. - -

Molecular dynamics

Key Type Default Allowed values Meaning Applies when Aliases
FMAX_ABORT float - 0.0.. MD force-safety abort threshold in eV/Angstrom. - -
SEED int - 0.. Random seed for reproducible MD. - DEFAULT_SEED

Output

Key Type Default Allowed values Meaning Applies when Aliases
WRITE_RESOLVED_CONFIG bool .TRUE. - Write resolved_config.json before a calculation. - -
WRITE_FORCES bool .TRUE. - Write forces to OUTCAR. - -
WRITE_OUTCAR bool - - Write the VASP-like OUTCAR interoperability view. - -
WRITE_XDATCAR bool - - Write the VASP XDATCAR interoperability trajectory. - -
WRITE_STRESS bool .TRUE. - Write stress to OUTCAR. - -
WRITE_TRAJECTORY bool - - Write ASE trajectory. - -
WRITE_JSON bool - - Write JSON results. - -
OUTPUT_FORMAT str VASP - Output format (VASP). - -
OUTPUT_DIR str - - Output directory for the calculation. Used when the command line does not pass --output; the resolved config is written into the final job directory. Relative values resolve against the file that declares them (the INCAR folder for an INCAR), and mliport run defaults to /results when neither --output nor OUTPUT_DIR is given. - -
JOB_NAME str - - Optional job name. - -

General

Key Type Default Allowed values Meaning Applies when Aliases
STRICT_CONFIG bool .TRUE. - Treat unknown configuration keys as errors. - -

Molecular dynamics

CALC_TYPE = MD runs NVT or NPT MD. Thermostat-specific keys only apply to the thermostat named in the "Applies when" column; the NPT keys only apply to ENSEMBLE = NPT.

Calculation

Key Type Default Allowed values Meaning Applies when Aliases
CALC_TYPE str - sp, opt, md, neb, batch, confsearch Top-level calculation type (selects the runner). - CALCULATION

Model

Key Type Default Allowed values Meaning Applies when Aliases
MODEL_TYPE str - uma, fairchem, mace, dpa, grace, nep MLIP engine. - -
MODEL_PATH str required (REQUIRED) - Model checkpoint file or directory. A value without any path separator is a file name: it is resolved next to the declaring file first, then in /models (one level of subdirectories allowed). - -
TASK str - - UMA task (omat/omol/...) or bulk/molecule for other engines. - -
INFERENCE_MODE str default default, turbo UMA inference mode. Non-default modes are rejected for other engines. - -
DEFAULT_DTYPE str - float32, float64 MACE model dtype (accuracy-first default: float64). - -
HEAD str - - MACE or DeepMD model head/branch (multi-task models). - -
ACTIVATION_CHECKPOINTING bool - - GPU memory saving (UMA, overrides inference_mode preset). - -
GPU_MEMORY_GROWTH bool - - GRACE TensorFlow allocator grows on demand instead of claiming the whole visible GPU. - -
GPU_MEMORY_LIMIT_MB int - 1.. Hard GRACE TensorFlow logical-device memory limit in MiB. - -
NEIGHBOR_CACHE bool - - GRACE verlet-style neighbor-list cache (extended cutoff + exact re-filter). The complete periodic-image multiset matches a fresh search; pair ordering may differ. Default: True. - -
NEIGHBOR_SKIN float - 1e-12.. GRACE neighbor-cache skin in Å: the cached neighbor list is rebuilt when any atom moves more than skin/2. Larger values rebuild less often but filter longer lists. - -
D3 bool - - Add Grimme DFT-D3 dispersion (torch-dftd) on top of the model. Only for models trained on dispersion-free reference data. - -
D3_XC str - - Reference functional of the model's training data, selecting the D3 parameters (default: pbe). - -
D3_DAMPING str - zero, bj, zerom, bjm DFT-D3 damping function (default: bj, i.e. D3(BJ)). - -
DPA_RUNTIME str - current, legacy DPA runtime environment: 'current' (pinned DPA-4-capable deepmd-kit) or 'legacy' (.venv-dpa-legacy, deepmd-kit 3.0.0b3, for DPA-2 checkpoints such as DPA-SSE). No automatic fallback. - -
NEP_RUNTIME str - ase, gpumd NEP execution runtime: 'ase' (Calorine CPUNEP/GPUNEP calculator) or 'gpumd' (native gpumd process). Required for MODEL_TYPE=nep; mliport never infers it or falls back between the two. - -
NEP_GPUMD_COMMAND str - - Path to the gpumd executable for NEP_RUNTIME=gpumd (and for the GPUNEP ASE adapter). Launched as an argv list, never through a shell; its real path, version and SHA256 are recorded. - -
CHARGE int - -100..100 Total molecular charge stored in atoms.info. - -
SPIN int - 0..100 Molecular spin metadata; UMA omol uses spin multiplicity (2S+1). - SPIN_MULTIPLICITY

Device and resources

Key Type Default Allowed values Meaning Applies when Aliases
DEVICE str cpu - Compute device: cpu, cuda, gpu or cuda:N. - -
TORCH_NUM_THREADS int - 1.. Backend CPU intra-op threads (PyTorch for UMA/MACE/DPA .pt; TensorFlow for GRACE; legacy canonical field name). - CPU_THREADS

Molecular dynamics

Key Type Default Allowed values Meaning Applies when Aliases
FMAX_ABORT float - 0.0.. MD force-safety abort threshold in eV/Angstrom. - -
MD_ENSEMBLE str NVT NVT, NVE, NPT MD ensemble. - -
TEMPERATURE float 300.0 0.0.. Temperature in Kelvin. - -
PRESSURE_GPA float 0.0 - NPT hydrostatic target pressure in GPa (positive is compressive). ENSEMBLE = NPT PRESSURE
NPT_TDAMP float 100.0 0.0.. MTK thermostat damping time in fs (strictly positive). ENSEMBLE = NPT -
NPT_PDAMP float 1000.0 0.0.. MTK barostat damping time in fs (strictly positive). ENSEMBLE = NPT -
NPT_TCHAIN int 3 1.. MTK NPT only: number of Nose-Hoover chain variables in the thermostat; >= 1, default 3 ENSEMBLE = NPT -
NPT_PCHAIN int 3 1.. MTK NPT only: number of chain variables in the barostat; >= 1, default 3 ENSEMBLE = NPT -
NPT_TLOOP int 1 1.. MTK NPT only: number of Suzuki-Yoshida sub-steps per thermostat integration; >= 1, default 1 ENSEMBLE = NPT -
NPT_PLOOP int 1 1.. MTK NPT only: number of Suzuki-Yoshida sub-steps per barostat integration; >= 1, default 1 ENSEMBLE = NPT -
TIMESTEP float 1.0 0.0.. Time step in femtoseconds. - -
STEPS int 1000 0.. Number of production MD steps. - PRODUCTION_STEPS
EQUILIBRATION_STEPS int 0 0.. Number of same-ensemble equilibration steps before production. - EQUIL_STEPS
THERMOSTAT str LANGEVIN LANGEVIN, BUSSI, NHC NVT thermostat (Langevin, Bussi/CSVR, or Nose-Hoover chain). - -
FRICTION float 0.001 0.0.. Langevin friction coefficient (1/fs). THERMOSTAT = LANGEVIN -
BUSSI_TAU float 1000.0 0.0.. Bussi/CSVR coupling time in femtoseconds. THERMOSTAT = BUSSI -
NHC_TDAMP float 100.0 0.0.. Nose-Hoover-chain damping time in femtoseconds. THERMOSTAT = NHC -
NHC_TCHAIN int 3 1.. Number of Nose-Hoover thermostat variables. THERMOSTAT = NHC -
NHC_TLOOP int 1 1.. Nose-Hoover thermostat integration substeps. THERMOSTAT = NHC -
SAVE_INTERVAL int 10 1.. Interval for saving trajectory frames. - TRAJECTORY_INTERVAL
PRE_RELAX bool - - Pre-relax the structure before MD (legacy). - -
PRE_RELAX_STEPS int 50 0.. Maximum pre-relaxation steps (legacy). - -
PRE_RELAX_FMAX float 0.1 0.0.. Pre-relaxation force threshold (legacy). - -
PRE_RELAX_MODE str none none, positions, cell Pre-relaxation mode (plan section 13.2; Phase 3). - -
VELOCITY_POLICY str auto auto, initialize, preserve Velocity initialisation policy (plan section 13.5; Phase 3). - -
COM_POLICY str auto auto, none, initialize_only, constraint Explicit center-of-mass policy for MD. - -
SEED int - 0.. Random seed for reproducible MD. - DEFAULT_SEED

Output

Key Type Default Allowed values Meaning Applies when Aliases
WRITE_RESOLVED_CONFIG bool .TRUE. - Write resolved_config.json before a calculation. - -
WRITE_FORCES bool .TRUE. - Write forces to OUTCAR. - -
WRITE_OUTCAR bool - - Write the VASP-like OUTCAR interoperability view. - -
WRITE_XDATCAR bool - - Write the VASP XDATCAR interoperability trajectory. - -
WRITE_STRESS bool .TRUE. - Write stress to OUTCAR. - -
WRITE_TRAJECTORY bool - - Write ASE trajectory. - -
WRITE_JSON bool - - Write JSON results. - -
OUTPUT_FORMAT str VASP - Output format (VASP). - -
OUTPUT_DIR str - - Output directory for the calculation. Used when the command line does not pass --output; the resolved config is written into the final job directory. Relative values resolve against the file that declares them (the INCAR folder for an INCAR), and mliport run defaults to /results when neither --output nor OUTPUT_DIR is given. - -
JOB_NAME str - - Optional job name. - -

General

Key Type Default Allowed values Meaning Applies when Aliases
STRICT_CONFIG bool .TRUE. - Treat unknown configuration keys as errors. - -

NEB

CALCULATION = NEB (CALC_TYPE = NEB is the same key) runs a fixed-cell nudged elastic band between two endpoint structures. NEB refuses the output switches that would silently drop a mandatory record.

Calculation

Key Type Default Allowed values Meaning Applies when Aliases
CALC_TYPE str - sp, opt, md, neb, batch, confsearch Top-level calculation type (selects the runner). - CALCULATION

Model

Key Type Default Allowed values Meaning Applies when Aliases
MODEL_TYPE str - uma, fairchem, mace, dpa, grace, nep MLIP engine. - -
MODEL_PATH str required (REQUIRED) - Model checkpoint file or directory. A value without any path separator is a file name: it is resolved next to the declaring file first, then in /models (one level of subdirectories allowed). - -
TASK str - - UMA task (omat/omol/...) or bulk/molecule for other engines. - -
INFERENCE_MODE str default default, turbo UMA inference mode. Non-default modes are rejected for other engines. - -
DEFAULT_DTYPE str - float32, float64 MACE model dtype (accuracy-first default: float64). - -
HEAD str - - MACE or DeepMD model head/branch (multi-task models). - -
ACTIVATION_CHECKPOINTING bool - - GPU memory saving (UMA, overrides inference_mode preset). - -
GPU_MEMORY_GROWTH bool - - GRACE TensorFlow allocator grows on demand instead of claiming the whole visible GPU. - -
GPU_MEMORY_LIMIT_MB int - 1.. Hard GRACE TensorFlow logical-device memory limit in MiB. - -
NEIGHBOR_CACHE bool - - GRACE verlet-style neighbor-list cache (extended cutoff + exact re-filter). The complete periodic-image multiset matches a fresh search; pair ordering may differ. Default: True. - -
NEIGHBOR_SKIN float - 1e-12.. GRACE neighbor-cache skin in Å: the cached neighbor list is rebuilt when any atom moves more than skin/2. Larger values rebuild less often but filter longer lists. - -
D3 bool - - Add Grimme DFT-D3 dispersion (torch-dftd) on top of the model. Only for models trained on dispersion-free reference data. - -
D3_XC str - - Reference functional of the model's training data, selecting the D3 parameters (default: pbe). - -
D3_DAMPING str - zero, bj, zerom, bjm DFT-D3 damping function (default: bj, i.e. D3(BJ)). - -
DPA_RUNTIME str - current, legacy DPA runtime environment: 'current' (pinned DPA-4-capable deepmd-kit) or 'legacy' (.venv-dpa-legacy, deepmd-kit 3.0.0b3, for DPA-2 checkpoints such as DPA-SSE). No automatic fallback. - -
NEP_RUNTIME str - ase, gpumd NEP execution runtime: 'ase' (Calorine CPUNEP/GPUNEP calculator) or 'gpumd' (native gpumd process). Required for MODEL_TYPE=nep; mliport never infers it or falls back between the two. - -
NEP_GPUMD_COMMAND str - - Path to the gpumd executable for NEP_RUNTIME=gpumd (and for the GPUNEP ASE adapter). Launched as an argv list, never through a shell; its real path, version and SHA256 are recorded. - -
CHARGE int - -100..100 Total molecular charge stored in atoms.info. - -
SPIN int - 0..100 Molecular spin metadata; UMA omol uses spin multiplicity (2S+1). - SPIN_MULTIPLICITY

Device and resources

Key Type Default Allowed values Meaning Applies when Aliases
DEVICE str cpu - Compute device: cpu, cuda, gpu or cuda:N. - -
TORCH_NUM_THREADS int - 1.. Backend CPU intra-op threads (PyTorch for UMA/MACE/DPA .pt; TensorFlow for GRACE; legacy canonical field name). - CPU_THREADS

Optimization

Key Type Default Allowed values Meaning Applies when Aliases
FMAX float 0.03 0.0.. Force convergence threshold in eV/Angstrom. - -
MAX_STEPS int 1000 0.. Maximum optimization steps. - -

Molecular dynamics

Key Type Default Allowed values Meaning Applies when Aliases
FMAX_ABORT float - 0.0.. MD force-safety abort threshold in eV/Angstrom. - -
SEED int - 0.. Random seed for reproducible MD. - DEFAULT_SEED

NEB

Key Type Default Allowed values Meaning Applies when Aliases
NEB_INITIAL str required (REQUIRED) - Initial endpoint structure for NEB. - -
NEB_FINAL str required (REQUIRED) - Final endpoint structure for NEB. - -
NEB_IMAGES int 7 1.. Number of intermediate images; endpoints are additional. - IMAGES
CLIMB bool .FALSE. - Run a second climbing-image NEB stage. - NEB_CLIMB
NEB_METHOD str improvedtangent improvedtangent ASE NEB tangent method. - -
NEB_INTERPOLATION str linear linear, idpp Initial band interpolation method. - -
NEB_PATH_CONVENTION str mic mic, unwrapped Periodic path convention. - -
NEB_SPRING float 0.1 1e-12.. NEB spring constant in eV/Angstrom^2. - -
NEB_PRE_FMAX float 0.1 1e-12.. Ordinary pre-stage force criterion in eV/Angstrom. - -
NEB_PRE_MAX_STEPS int 300 0.. Maximum ordinary pre-stage optimizer steps. - -
NEB_MAXSTEP float 0.1 1e-12.. Maximum FIRE displacement per step in Angstrom. - -
NEB_ENDPOINT_POLICY str validate validate, relax Validate or relax endpoint geometries before NEB. - -
NEB_ENDPOINT_FMAX float 0.02 1e-12.. Endpoint convergence threshold in eV/Angstrom. - -
NEB_ENDPOINT_STEPS int 500 1.. Maximum endpoint relaxation steps. - -
IDPP_FMAX float 0.1 1e-12.. IDPP interpolation force criterion. - -
IDPP_STEPS int 200 1.. Maximum IDPP interpolation steps. - -
IDPP_MIC bool .TRUE. - Use minimum-image distances in IDPP. - -
NEB_MIN_DISTANCE float 0.5 1e-12.. Minimum allowed interatomic distance in Angstrom. - -
NEB_CHECKPOINT_INTERVAL int 10 1.. Trusted optimizer-state interval between full-band checkpoints. - -

Output

Key Type Default Allowed values Meaning Applies when Aliases
WRITE_RESOLVED_CONFIG bool .TRUE. - Write resolved_config.json before a calculation. - -
WRITE_FORCES bool .TRUE. - Write forces to OUTCAR. - -
WRITE_OUTCAR bool - - Write the VASP-like OUTCAR interoperability view. - -
WRITE_XDATCAR bool - - Write the VASP XDATCAR interoperability trajectory. - -
WRITE_STRESS bool .TRUE. - Write stress to OUTCAR. - -
WRITE_TRAJECTORY bool - - Write ASE trajectory. - -
WRITE_JSON bool - - Write JSON results. - -
OUTPUT_FORMAT str VASP - Output format (VASP). - -
OUTPUT_DIR str - - Output directory for the calculation. Used when the command line does not pass --output; the resolved config is written into the final job directory. Relative values resolve against the file that declares them (the INCAR folder for an INCAR), and mliport run defaults to /results when neither --output nor OUTPUT_DIR is given. - -
JOB_NAME str - - Optional job name. - -

General

Key Type Default Allowed values Meaning Applies when Aliases
STRICT_CONFIG bool .TRUE. - Treat unknown configuration keys as errors. - -

CALC_TYPE = CONFSEARCH runs the MD-based configuration search. mliport run executes it through the same code path as mliport confsearch run.

Calculation

Key Type Default Allowed values Meaning Applies when Aliases
CALC_TYPE str - sp, opt, md, neb, batch, confsearch Top-level calculation type (selects the runner). - CALCULATION

Model

Key Type Default Allowed values Meaning Applies when Aliases
MODEL_TYPE str - uma, fairchem, mace, dpa, grace, nep MLIP engine. - -
MODEL_PATH str required (REQUIRED) - Model checkpoint file or directory. A value without any path separator is a file name: it is resolved next to the declaring file first, then in /models (one level of subdirectories allowed). - -
TASK str - - UMA task (omat/omol/...) or bulk/molecule for other engines. - -
INFERENCE_MODE str - default, turbo UMA inference mode. Non-default modes are rejected for other engines. - -
DEFAULT_DTYPE str - float32, float64 MACE model dtype (accuracy-first default: float64). - -
HEAD str - - MACE or DeepMD model head/branch (multi-task models). - -
ACTIVATION_CHECKPOINTING bool - - GPU memory saving (UMA, overrides inference_mode preset). - -
GPU_MEMORY_GROWTH bool - - GRACE TensorFlow allocator grows on demand instead of claiming the whole visible GPU. - -
GPU_MEMORY_LIMIT_MB int - 1.. Hard GRACE TensorFlow logical-device memory limit in MiB. - -
NEIGHBOR_CACHE bool - - GRACE verlet-style neighbor-list cache (extended cutoff + exact re-filter). The complete periodic-image multiset matches a fresh search; pair ordering may differ. Default: True. - -
NEIGHBOR_SKIN float - 1e-12.. GRACE neighbor-cache skin in Å: the cached neighbor list is rebuilt when any atom moves more than skin/2. Larger values rebuild less often but filter longer lists. - -
D3 bool - - Add Grimme DFT-D3 dispersion (torch-dftd) on top of the model. Only for models trained on dispersion-free reference data. - -
D3_XC str - - Reference functional of the model's training data, selecting the D3 parameters (default: pbe). - -
D3_DAMPING str - zero, bj, zerom, bjm DFT-D3 damping function (default: bj, i.e. D3(BJ)). - -
DPA_RUNTIME str - current, legacy DPA runtime environment: 'current' (pinned DPA-4-capable deepmd-kit) or 'legacy' (.venv-dpa-legacy, deepmd-kit 3.0.0b3, for DPA-2 checkpoints such as DPA-SSE). No automatic fallback. - -
NEP_RUNTIME str - ase, gpumd NEP execution runtime: 'ase' (Calorine CPUNEP/GPUNEP calculator) or 'gpumd' (native gpumd process). Required for MODEL_TYPE=nep; mliport never infers it or falls back between the two. - -
NEP_GPUMD_COMMAND str - - Path to the gpumd executable for NEP_RUNTIME=gpumd (and for the GPUNEP ASE adapter). Launched as an argv list, never through a shell; its real path, version and SHA256 are recorded. - -
CHARGE int - -100..100 Total molecular charge stored in atoms.info. - -
SPIN int - 0..100 Molecular spin metadata; UMA omol uses spin multiplicity (2S+1). - SPIN_MULTIPLICITY

Device and resources

Key Type Default Allowed values Meaning Applies when Aliases
DEVICE str cpu - Compute device: cpu, cuda, gpu or cuda:N. - -
TORCH_NUM_THREADS int - 1.. Backend CPU intra-op threads (PyTorch for UMA/MACE/DPA .pt; TensorFlow for GRACE; legacy canonical field name). - CPU_THREADS

Molecular dynamics

Key Type Default Allowed values Meaning Applies when Aliases
FMAX_ABORT float - 0.0.. MD force-safety abort threshold in eV/Angstrom. - -
TIMESTEP float 0.5 0.0.. Time step in femtoseconds. - -
FRICTION float 0.01 0.0.. Langevin friction coefficient (1/fs). - -
SEED int - 0.. Random seed for reproducible MD. - DEFAULT_SEED

Configuration search

Key Type Default Allowed values Meaning Applies when Aliases
CS_MODE str adsorbate adsorbate Search mode; v1 accepts only 'adsorbate' (bulk_mobile, melt_quench, interface, multi_adsorbate are rejected) - -
CS_STARTS str required (REQUIRED) - Mode A: start structures (files or directories), one independent replica each; needs cs_adsorbate_indices - -
CS_ADSORBATE_INDICES str required (REQUIRED) - Mode A: 0-based adsorbate atom indices such as 300-341 or 0,5,7-9; must be one connected covalent molecule - -
CS_SLAB str slab.vasp (example) - Mode B: slab structure file, exactly one frame; conformers from cs_molecule are placed above it - -
CS_MOLECULE str molecule.xyz (example) - Mode B: molecule conformer file; one conformer is drawn at random for each placed start - -
CS_N_STARTS int 10 1.. Mode B: number of placed start structures to build, each becoming one independent replica; minimum 1 - -
CS_FIX_BELOW float - - Extend the input FixAtoms to slab atoms within this height in Angstrom of the lowest slab atom; unset = off - -
CS_MD_MOBILE str adsorbate adsorbate, adsorbate_and_top, all_free Which atoms are mobile in MD: adsorbate (default), adsorbate_and_top or all_free (every atom not fixed) - -
CS_TOP_DEPTH float 2.0 - Depth in Angstrom below the slab top of the extra mobile atoms in cs_md_mobile=adsorbate_and_top; default 2.0 - -
CS_TEMPERATURE_HIGH float 500.0 - Hold temperature in K of every heat-hold-cool cycle and of the initial velocities; default 500 - -
CS_TEMPERATURE_LOW float 300.0 - End temperature in K of the linear cooling ramp in each cycle, must be <= cs_temperature_high; default 300 - -
CS_HOLD_PS float 2.0 - Duration in ps of the hold phase at cs_temperature_high in each cycle; must be a whole number of timesteps - -
CS_COOL_PS float 2.0 - Duration in ps of the linear cooling ramp from cs_temperature_high to cs_temperature_low in each cycle - -
CS_SNAPSHOT_INTERVAL_PS float 0.5 - Spacing in ps between hold-phase snapshots that get quenched; must be < cs_hold_ps, 0 disables snapshots - -
CS_RAMP_INTERVAL int 10 1.. Number of MD steps between thermostat set-point updates during the linear cooling ramp; default 10 - -
CS_CYCLES int 5 0.. Number of heat-hold-cool cycles run per replica after its start quench; 0 leaves only the start quenches - -
CS_SAVE_INTERVAL int 100 1.. Write one raw MD frame every this many MD steps (default 100); quench snapshots and cycle ends are always kept - -
CS_ACCEPT_TEMPERATURE float 500.0 - Temperature in K of the Metropolis test that accepts or rejects a replica's end-of-cycle minimum; default 500 - -
CS_QUENCH_FMAX float 0.05 - FIRE force convergence threshold in eV/A for every snapshot and cycle-end quench; default 0.05 - -
CS_QUENCH_STEPS int 2000 1.. Maximum FIRE steps per quench; exceeding it marks the minimum quench_not_converged and unranked; default 2000 - -
CS_BOND_SCALE float 1.2 - Bond test in the covalent graphs: bonded when closer than this multiple of the covalent-radii sum; default 1.2 - -
CS_BOND_BREAK_SCALE float 1.5 - MD aborts if an adsorbate bond exceeds this multiple of the covalent-radii sum; must be > cs_bond_scale (1.5) - -
CS_IONIC_SPECIES str Li,Na,K,Rb,Cs,Mg,Ca - Element symbols excluded from every covalent graph (ionic hopping is no reaction); default Li,Na,K,Rb,Cs,Mg,Ca - -
CS_REACTION_POLICY str exclude exclude, include Whether minima whose covalent bonds changed are ranked and exported: exclude (default) or include - -
CS_DESORPTION_DISTANCE float 5.0 - Adsorbate-slab distance in Angstrom above which a minimum is labelled desorbed and dropped; default 5.0 - -
CS_MIN_IMAGE_DISTANCE float 4.0 - Adsorbate in-plane self-image distance in Angstrom below which a minimum is image_contact; default 4.0 - -
CS_MIN_VACUUM_GAP float 8.0 - Gap in Angstrom from the adsorbate to the slab image one cell along c; smaller marks image_contact (8.0) - -
CS_PLACE_CONTACT_SCALE float 1.5 - Mode B: target ratio of the closest adsorbate-slab pair to the covalent-radii sum (unbonded); default 1.5 - -
CS_PLACE_ANCHOR str - - Mode B: 0-based anchor atom indices in the molecule file; the lowest atom is drawn from them and placed bonded - -
CS_PLACE_ANCHOR_SCALE float 1.1 - Mode B with cs_place_anchor: anchored pair ratio to the covalent-radii sum, i.e. already bonded; default 1.1 - -
CS_PLACE_OTHER_MIN_SCALE float 1.3 - Mode B with cs_place_anchor: lower bound for every other adsorbate-slab pair in radii-sum units; default 1.3 - -
CS_PLACE_ANCHOR_TARGETS str - - Mode B with cs_place_anchor: slab elements the closest slab atom must belong to (e.g. an ionic species) - -
CS_PLACE_MAX_ATTEMPTS int 1000 1.. Mode B: random placement attempts tried per start before the run fails; default 1000 - -
CS_DEDUP_ENERGY_TOL float 0.01 - Energy window in eV within which two minima may be merged as the same minimum; default 0.010 - -
CS_DEDUP_GEOM_TOL float 0.2 - Per-atom displacement in Angstrom below which two minima count as the same geometry (rule G); default 0.20 - -
CS_DEDUP_FP_TOL float 0.0004 - Fingerprint distance below which two minima merge (rule D, catches symmetry equivalents); default 0.0004 - -
CS_FP_RCUT float 8.0 - Cutoff radius in Angstrom of the adsorbate Oganov-Valle fingerprint used for dedup; default 8.0 - -
CS_FP_BINWIDTH float 0.02 - Fingerprint histogram bin width in Angstrom, at most half of cs_fp_sigma; default 0.02 - -
CS_FP_SIGMA float 0.2 - Gaussian smearing width in Angstrom applied to the fingerprint histogram bins; default 0.20 - -
CS_TOP_K int 10 1.. Maximum number of ranked minima exported as candidates for DFT re-ranking; default 10 - -
CS_EXPORT_WINDOW float 0.5 - Only minima within this energy window in eV above the best one are refined and exported; default 0.5 - -
CS_REFINE bool .TRUE. - Re-refine the exported candidates with a tighter force threshold before ranking and export; default true - -
CS_REFINE_FMAX float 0.02 - Force threshold in eV/A for the final refinement, must be <= cs_quench_fmax; default 0.02 - -
CS_REFINE_STEPS int 3000 1.. Maximum number of FIRE steps in the final refinement of each exported minimum; default 3000 - -
CS_ALLOW_DIRECT_FORCES bool .FALSE. - Allow a model that reports direct_forces=true, in which case the ranking is not meaningful; default false - -

Output

Key Type Default Allowed values Meaning Applies when Aliases
WRITE_RESOLVED_CONFIG bool .TRUE. - Write resolved_config.json before a calculation. - -
WRITE_FORCES bool .TRUE. - Write forces to OUTCAR. - -
WRITE_OUTCAR bool - - Write the VASP-like OUTCAR interoperability view. - -
WRITE_XDATCAR bool - - Write the VASP XDATCAR interoperability trajectory. - -
WRITE_STRESS bool .TRUE. - Write stress to OUTCAR. - -
WRITE_TRAJECTORY bool - - Write ASE trajectory. - -
WRITE_JSON bool - - Write JSON results. - -
OUTPUT_FORMAT str VASP - Output format (VASP). - -
OUTPUT_DIR str - - Output directory for the calculation. Used when the command line does not pass --output; the resolved config is written into the final job directory. Relative values resolve against the file that declares them (the INCAR folder for an INCAR), and mliport run defaults to /results when neither --output nor OUTPUT_DIR is given. - -
JOB_NAME str - - Optional job name. - -

General

Key Type Default Allowed values Meaning Applies when Aliases
STRICT_CONFIG bool .TRUE. - Treat unknown configuration keys as errors. - -

Batch

mliport batch sweeps a directory of structures instead of one INCAR run, so these keys are read from settings.ini or the command line; an INCAR.mliport whose CALC_TYPE = BATCH is refused by mliport run with a pointer to the batch subcommand.

Calculation

Key Type Default Allowed values Meaning Applies when Aliases
CALC_TYPE str - sp, opt, md, neb, batch, confsearch Top-level calculation type (selects the runner). - CALCULATION

Model

Key Type Default Allowed values Meaning Applies when Aliases
MODEL_TYPE str - uma, fairchem, mace, dpa, grace, nep MLIP engine. - -
MODEL_PATH str required (REQUIRED) - Model checkpoint file or directory. A value without any path separator is a file name: it is resolved next to the declaring file first, then in /models (one level of subdirectories allowed). - -
TASK str - - UMA task (omat/omol/...) or bulk/molecule for other engines. - -
INFERENCE_MODE str - default, turbo UMA inference mode. Non-default modes are rejected for other engines. - -
DEFAULT_DTYPE str - float32, float64 MACE model dtype (accuracy-first default: float64). - -
HEAD str - - MACE or DeepMD model head/branch (multi-task models). - -
ACTIVATION_CHECKPOINTING bool - - GPU memory saving (UMA, overrides inference_mode preset). - -
GPU_MEMORY_GROWTH bool - - GRACE TensorFlow allocator grows on demand instead of claiming the whole visible GPU. - -
GPU_MEMORY_LIMIT_MB int - 1.. Hard GRACE TensorFlow logical-device memory limit in MiB. - -
NEIGHBOR_CACHE bool - - GRACE verlet-style neighbor-list cache (extended cutoff + exact re-filter). The complete periodic-image multiset matches a fresh search; pair ordering may differ. Default: True. - -
NEIGHBOR_SKIN float - 1e-12.. GRACE neighbor-cache skin in Å: the cached neighbor list is rebuilt when any atom moves more than skin/2. Larger values rebuild less often but filter longer lists. - -
D3 bool - - Add Grimme DFT-D3 dispersion (torch-dftd) on top of the model. Only for models trained on dispersion-free reference data. - -
D3_XC str - - Reference functional of the model's training data, selecting the D3 parameters (default: pbe). - -
D3_DAMPING str - zero, bj, zerom, bjm DFT-D3 damping function (default: bj, i.e. D3(BJ)). - -
DPA_RUNTIME str - current, legacy DPA runtime environment: 'current' (pinned DPA-4-capable deepmd-kit) or 'legacy' (.venv-dpa-legacy, deepmd-kit 3.0.0b3, for DPA-2 checkpoints such as DPA-SSE). No automatic fallback. - -
NEP_RUNTIME str - ase, gpumd NEP execution runtime: 'ase' (Calorine CPUNEP/GPUNEP calculator) or 'gpumd' (native gpumd process). Required for MODEL_TYPE=nep; mliport never infers it or falls back between the two. - -
NEP_GPUMD_COMMAND str - - Path to the gpumd executable for NEP_RUNTIME=gpumd (and for the GPUNEP ASE adapter). Launched as an argv list, never through a shell; its real path, version and SHA256 are recorded. - -
CHARGE int - -100..100 Total molecular charge stored in atoms.info. - -
SPIN int - 0..100 Molecular spin metadata; UMA omol uses spin multiplicity (2S+1). - SPIN_MULTIPLICITY

Device and resources

Key Type Default Allowed values Meaning Applies when Aliases
DEVICE str cpu - Compute device: cpu, cuda, gpu or cuda:N. - -
TORCH_NUM_THREADS int - 1.. Backend CPU intra-op threads (PyTorch for UMA/MACE/DPA .pt; TensorFlow for GRACE; legacy canonical field name). - CPU_THREADS

Optimization

Key Type Default Allowed values Meaning Applies when Aliases
OPT_ALGO str - FIRE, BFGS, LBFGS Geometry optimization algorithm. - -
FMAX float - 0.0.. Force convergence threshold in eV/Angstrom. - -
MAX_STEPS int - 0.. Maximum optimization steps. - -
CELL_OPT bool - - Optimize cell parameters (requires stress support). - -
FIX_SYMMETRY bool - - Preserve crystal symmetry during optimization. - -

Molecular dynamics

Key Type Default Allowed values Meaning Applies when Aliases
FMAX_ABORT float - 0.0.. MD force-safety abort threshold in eV/Angstrom. - -
SEED int - 0.. Random seed for reproducible MD. - DEFAULT_SEED

Output

Key Type Default Allowed values Meaning Applies when Aliases
WRITE_RESOLVED_CONFIG bool .TRUE. - Write resolved_config.json before a calculation. - -
WRITE_FORCES bool .TRUE. - Write forces to OUTCAR. - -
WRITE_OUTCAR bool - - Write the VASP-like OUTCAR interoperability view. - -
WRITE_XDATCAR bool - - Write the VASP XDATCAR interoperability trajectory. - -
WRITE_STRESS bool .TRUE. - Write stress to OUTCAR. - -
WRITE_TRAJECTORY bool - - Write ASE trajectory. - -
WRITE_JSON bool - - Write JSON results. - -
OUTPUT_FORMAT str VASP - Output format (VASP). - -
OUTPUT_DIR str - - Output directory for the calculation. Used when the command line does not pass --output; the resolved config is written into the final job directory. Relative values resolve against the file that declares them (the INCAR folder for an INCAR), and mliport run defaults to /results when neither --output nor OUTPUT_DIR is given. - -
JOB_NAME str - - Optional job name. - -

Batch

Key Type Default Allowed values Meaning Applies when Aliases
SUB_CALC_TYPE str - sp, opt Sub-calculation type for a sweep batch (sp/opt). - -
PATTERN str - - Glob pattern for batch structure discovery. - -

General

Key Type Default Allowed values Meaning Applies when Aliases
STRICT_CONFIG bool .TRUE. - Treat unknown configuration keys as errors. - -