+------------------------------------------------------------+
| This is experimental code with lots of bugs intended for |
| internal use only. Will be ready as soon as we publish the |
| methodology. |
+------------------------------------------------------------+
Automate the modeling of thermodynamic properties in organic molecular crystals using machine-learning interatomic potentials (MLIPs). The general computational scheme facilitated by this program involves the following steps:
+-------------------------------------------------------------+
| Generate training set configurations for a periodic system |
| using a baseline MLIP (e.g., MACE). |
+--------------------------+----------------------------------+
|
v
+--------------------------+----------------------------------+
| Extract finite subsystems composed of one or more molecules.|
+--------------------------+----------------------------------+
|
v
+--------------------------+----------------------------------+
| Use these finite subsystems to generate high-accuracy data |
| points using quantum chemical wave function methods. |
+--------------------------+----------------------------------+
|
v
+--------------------------+----------------------------------+
| Perform training of a delta-learning layer on top of the |
| baseline model. |
+--------------------------+----------------------------------+
|
v
+--------------------------+----------------------------------+
| Carry out the final calculation (e.g., quasi-harmonic |
| thermodynamics or molecular dynamics) using the newly |
| trained model. |
+-------------------------------------------------------------+
The program integrates several scientific codes into a unified workflow, including:
- MACE: Primary MLIP model for energies and forces.
- phonopy: Phonon calculations and vibrational properties.
- pymatgen: Crystal structure analysis and manipulation.
- ASE: Geometry relaxation and molecular dynamics simulations.
- PySCF and GPU4PySCF: Mean-field electronic structure calculations.
- MRCC and beyond-rpa: High-fidelity data points from correlated wave-function theory.