Version: Next

Silicon Example (Foundation-Model Sampling)

This example demonstrates active learning for silicon using a foundation model (EquiformerV2) and DIRECT sampling. PWact/examples/ provides several combinations; see PWact/examples/README.md for details. Mcloud users can find the examples under /share/public/PWMLFF_test_data/pwact_examples/25-pwact-demo. The [PWact/examples/si_pwmatgaussion_bigmodel_direct] example illustrates the parameters and scripts for foundation-model and DIRECT workflows. For constructing an initial training set with init_bulk, see examples/si_pwmatgaussion_bigmodel_direct/init_bulk_bigmodel:

Step 1. Relax the structure with PWmat using a Gaussian basis.
Step 2. Run molecular dynamics with a foundation model (SevenNet).
Step 3. Apply DIRECT sampling to the MD trajectories, remove similar structures, and use the selected structures for subsequent active learning.

The Gaussian-basis PWmat settings in this directory are intended only for quick workflow testing. The generated data are not suitable for scientific use.

Prepare Pretraining Data with init_bulk

Command: Enter examples/si_pwmatgaussion_bigmodel_direct/init_bulk_bigmodel. The directory provides an Mcloud job script, or you can run:

pwact init_bulk init_param.json resource.json

init_bulk Directory Structure

The init_bulk directory resembles that of the si_pwmat example, with an additional bigmodel directory:

.
├── datapath.txt
├── init_config_0/
├── init_config_1/
└── bigmodel/
    ├── 0-bigmodel.job
    ├── 0-tag.bigmodel.success
    ├── 1-bigmodel.job
    ├── 1-tag.bigmodel.success
    ├── ...
    ├── direct/
    │   ├── 0-direct.job
    │   ├── 0-tag.direct.success
    │   ├── candidate.json
    │   ├── candidate.xyz
    │   ├── Cov_score.png
    │   ├── direct.py
    │   ├── PCA_direct.png
    │   ├── PCA_variance.png
    │   ├── PWdata
    │   ├── select_idx.dat
    │   ├── select.xyz
    │   └── tag.direct.success
    ├── init_config_0/
    │   ├── 0.95_scale/
    │   │   ├── 3_bigmodel/
    │   │   │   ├── npt.log
    │   │   │   ├── POSCAR
    │   │   │   ├── sevennet_md.py
    │   │   │   ├── tag.bigmodel.success
    │   │   │   ├── tmp.traj
    │   │   │   └── traj.xyz
    │   │   ├── 4_bigmodel/
    │   │   └── 5_bigmodel/
    │   ├── 0.9_scale/
    │   └── 1.0_scale/
    └── init_config_*/

*-bigmodel.job and *-tag.bigmodel.success are Slurm scripts for foundation-model MD and success-marker files.
init_config_0, init_config_1, and so on are working directories for foundation-model MD.
direct is the working directory for structure selection with DIRECT.

Integrate Foundation-Model MD into init_bulk

Foundation-model MD uses a user-defined interface and requires the corresponding runtime environment. See sevennet_md.py in the example.

As illustrated by bigmodel/init_config_0/0.95_scale/3_bigmodel/, when PWact detects a bigmodel setting during init_bulk, it creates a working directory for each structure, writes the structure as a VASP/POSCAR file, copies the user-provided MD interface script such as sevennet_md.py, generates a Slurm script from the explore settings in resource.json, and submits the job.

The interface script must read POSCAR, run MD, and convert the trajectory to an extxyz file named traj.xyz. PWact detects this file and uses it in subsequent processing.

Integrate DIRECT Sampling into init_bulk

DIRECT sampling uses a user-defined interface and requires a configured DIRECT environment. See direct.py in the example.

After foundation-model MD produces multiple trajectories, PWact detects and merges them into bigmodel/direct/candidate.xyz. It copies the DIRECT interface script specified by direct_input, such as direct.py, into that directory, generates a Slurm script from the direct settings in resource.json, and submits the job.

The script must read candidate.xyz and generate select_idx.dat and select.xyz. select_idx.dat stores the indices of selected structures in candidate.xyz, while select.xyz stores the selected structures themselves. PWact reads both files for subsequent steps.

Active Learning with run

This example uses the pretraining data and perturbed structures from init_bulk for active learning at 500 K, 800 K, and 1100 K.

Command: After init_bulk completes, enter examples/si_pwmatgaussion_bigmodel_direct/run_iter_direct_bigmodel:

pwact run param.json resource.json

Active-Learning Directory

The directory structure resembles that of the si_pwmat example.

train Directory

The train directory is identical to that in the si_pwmat example.

explore Directory

In addition to md and select, explore contains a bigmodel subdirectory whose contents match those described for DIRECT sampling in init_bulk.

The md and select subdirectories are identical to those in the si_pwmat example.

label Directory

When bigmodel_script is configured for foundation-model labeling, a bigmodel subdirectory is added under label:

bigmodel/
├── 0-bigmodel.job
├── eqv2_label.py
├── select.xyz
└── train.xyz

eqv2_label.py is the user-specified bigmodel_script; PWact copies it into bigmodel.
select.xyz is the extxyz input containing structures to label.
train.xyz is the labeled extxyz output. It preserves the structure order of select.xyz and adds energies and forces.
0-bigmodel.job is the Slurm script generated from the DFT settings in resource.json.

Integrate Foundation-Model Labeling into run

Foundation-model labeling uses a user-defined interface and requires the corresponding runtime environment. See eqv2_label.py in the example.

After multi-model deviation selection—or additional DIRECT selection—during explore, PWact detects the selected structures, merges them into bigmodel/select.xyz, copies the interface script specified by bigmodel_script, such as eqv2_label.py, generates a Slurm script from the DFT settings in resource.json, and submits the job.

The script must read select.xyz and generate train.xyz, adding energy and force information. PWact reads this file for subsequent steps.

result

result collects the labeled dataset. When data_format is extxyz, the result is train.xyz; when it is pwmlff/npy, the result is a PWdata directory containing the dataset files.