mfi-M3GNet implementation is added (#77)

* mfi-M3GNet implementation is added

* adding the description of stress_weight in _dynamics.py

* fix the version of pip in requirements.txt

Avoid update of pip version for installing tensorflow==2.11.0.

* Update requirements.txt

* remove update of pip version in linting.yml

* remove upgrade of pip version in testing.yml

* Try to upgrade the version of Tensorflow

* fix Too many positional arguments for method call in _converters.py

* further fix Too many positional arguments for method call in _converters.py

* fix Possibly using variable 'mgb_val' before assignment in _property.py

* fix Possibly using variable 'mgb_val' before assignment in _potential.py

* fix black

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fixed nbqa-flake8 in jupyter notebook

* fix black

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* try to upgrade pymatgen

* fix black

* fix bug in spherical harmonic function

* fix the sympy version

* change back to _conjugate in _matg.py

* try to fix it by adding .ref()

* skip the test of isolated atom for now

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* skip the test of isolated atom for now again

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* put back the pip upgrade

* correct the upgrade of pip in testing.yaml

* added element_refs in _dynamics.py

* merge the conflicts in test_model.py

* fix pylint

* fix the import for AtomRef

* fix pytest

* fix pytest again

* fix pytest again

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Shyue Ping Ong <[email protected]>

examples/Relaxation of LiFePO4.ipynb

@@ -105,7 +105,7 @@
    ],
    "source": [
     "relaxer = Relaxer()\n",
-    "relax_results: dict\n",
+    "relax_results: dict = {}\n",
     "%time relax_results = relaxer.relax(lfp_strained)\n",
     "relaxed_struct = relax_results[\"final_structure\"]"
    ]
@@ -394,7 +394,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.9.13"
+   "version": "3.10.9"
   },
   "vscode": {
    "interpreter": {

m3gnet/__init__.py

@@ -1,4 +1,5 @@
 """
 The M3GNet framework package
 """
+
 __version__ = "0.2.4"

m3gnet/callbacks.py

@@ -1,6 +1,7 @@
 """
 Callback functions
 """
+
 import os
 from typing import Dict
 

m3gnet/config.py

@@ -1,4 +1,5 @@
 """Data types"""
+
 import numpy as np
 import tensorflow as tf
 

m3gnet/graph/_batch.py

@@ -1,6 +1,7 @@
 """
 Collate material graphs
 """
+
 from typing import AnyStr, List, Optional, Tuple, Union, overload
 
 import numpy as np
@@ -144,12 +145,12 @@ def _concatenate(list_of_arrays: List, name: AnyStr) -> Optional[np.ndarray]:
 
 @overload
 def assemble_material_graph(graphs: List[MaterialGraph]) -> MaterialGraph:
-    ...
+    raise NotImplementedError
 
 
 @overload
 def assemble_material_graph(graphs: List[List]) -> List:
-    ...
+    raise NotImplementedError
 
 
 def assemble_material_graph(graphs):

m3gnet/graph/_converters.py

@@ -1,6 +1,7 @@
 """
 Classes to convert a structure into a graph
 """
+
 import logging
 from abc import abstractmethod
 from typing import Dict, List, Optional
@@ -77,7 +78,7 @@ def get_states(self, structure: StructureOrMolecule):
         return states
 
     @abstractmethod
-    def convert(self, structure: StructureOrMolecule, **kwargs) -> MaterialGraph:
+    def convert(self, structure: StructureOrMolecule, state_attr=None, *args, **kwargs) -> MaterialGraph:
         """
         Convert the structure into a graph
         Args:
@@ -98,17 +99,18 @@ def convert_many(self, structures: List[StructureOrMolecule], **kwargs) -> Mater
         graphs = [self.convert(structure, **kwargs) for structure in structures]
         return assemble_material_graph(graphs)
 
-    def __call__(self, structure: StructureOrMolecule, *args, **kwargs) -> MaterialGraph:
+    def __call__(self, structure: StructureOrMolecule, state_attr=None, *args, **kwargs) -> MaterialGraph:
         """
         A thin wrapper for calling `convert` method
         Args:
             structure:
+            state_attr:
             *args:
             **kwargs:
         Returns:
 
         """
-        return self.convert(structure)
+        return self.convert(structure, state_attr, *args, **kwargs)
 
 
 @register
@@ -145,11 +147,12 @@ def __init__(
 
         super().__init__(**kwargs)
 
-    def convert(self, structure: StructureOrMolecule, **kwargs) -> MaterialGraph:
+    def convert(self, structure: StructureOrMolecule, state_attr=None, *args, **kwargs) -> MaterialGraph:
         """
         Convert the structure into graph
         Args:
             structure: Structure or Molecule
+            state_attr: Global state attribute (e.g.Fidelity of data)
         Returns
             MaterialGraph
         """
@@ -158,7 +161,7 @@ def convert(self, structure: StructureOrMolecule, **kwargs) -> MaterialGraph:
             atom_positions = np.asarray(structure.get_positions(), dtype=DataType.np_float)
         else:
             atom_positions = np.array(structure.cart_coords, dtype=DataType.np_float)
-        state_attributes = self.get_states(structure)
+        state_attributes = state_attr if state_attr is not None else self.get_states(structure)
 
         sender_indices, receiver_indices, images, distances = get_fixed_radius_bonding(structure, self.cutoff)
 

m3gnet/layers/__init__.py

@@ -1,6 +1,7 @@
 """
 Graph layers
 """
+
 from ._aggregate import AtomReduceState
 from ._atom import AtomNetwork, GatedAtomUpdate
 from ._atom_ref import AtomRef, BaseAtomRef

m3gnet/layers/_aggregate.py

@@ -1,4 +1,5 @@
 """Aggregate classes. Aggregating happens when bond attributes """
+
 from typing import Callable, List, Union
 
 import tensorflow as tf

m3gnet/layers/_atom_ref.py

@@ -60,7 +60,12 @@ def __init__(
         if property_per_element is None:
             self.property_per_element = np.zeros(shape=(max_z + 1,))
         else:
-            self.property_per_element = np.array(property_per_element).ravel()
+            self.property_per_element = np.array(property_per_element)
+            if self.property_per_element.ndim > 1:
+                self.n_state = self.property_per_element.shape[0]
+            else:
+                self.property_per_element = np.array(property_per_element).ravel()
+                self.n_state = 1
         self.max_z = max_z
 
     def _get_feature_matrix(self, structs_or_graphs):
@@ -144,7 +149,11 @@ def call(self, graph: List, **kwargs):
         Returns:
         """
         atomic_numbers = graph[Index.ATOMS][:, 0]
-        atom_energies = tf.gather(tf.cast(self.property_per_element, DataType.tf_float), atomic_numbers)
+        if self.n_state == 1:
+            atom_energies = tf.gather(tf.cast(self.property_per_element, DataType.tf_float), atomic_numbers)
+        else:
+            state_property_per_element = self.property_per_element[:, int(graph[Index.STATES])]
+            atom_energies = tf.gather(tf.cast(state_property_per_element, DataType.tf_float), atomic_numbers)
         res = tf.math.segment_sum(atom_energies, get_segment_indices_from_n(graph[Index.N_ATOMS]))
         return tf.reshape(res, (-1, 1))
 

m3gnet/layers/_base.py

@@ -1,6 +1,7 @@
 """
 Base layer classes
 """
+
 from typing import Callable, List, Optional
 
 import tensorflow as tf

m3gnet/layers/_core.py

@@ -1,6 +1,7 @@
 """
 Core layers provide basic operations, e.g., MLP
 """
+
 from typing import Dict, List, Union
 
 import tensorflow as tf

m3gnet/layers/_gn.py

@@ -1,6 +1,7 @@
 """
 Materials Graph Network
 """
+
 from copy import deepcopy
 from typing import List, Optional
 

m3gnet/layers/_readout.py

@@ -1,6 +1,7 @@
 """
 Readout compress a graph into a vector
 """
+
 from typing import List, Optional
 
 import tensorflow as tf
@@ -109,15 +110,17 @@ class ReduceReadOut(ReadOut):
     This could be summing up the atoms or bonds, or taking the mean, etc.
     """
 
-    def __init__(self, method: str = "mean", field="atoms", **kwargs):
+    def __init__(self, method: str = "mean", field="atoms", output_latent_feats: bool = False, **kwargs):
         """
         Args:
             method (str): method for the reduction
             field (str): the field of MaterialGraph to perform the reduction
+            output_latent_feats (bool): whether output latent atomic features
             **kwargs:
         """
         self.method = method
         self.field = field
+        self.output_latent_feats = output_latent_feats
         super().__init__(**kwargs)
         self.method_func = METHOD_MAPPING.get(method)
 
@@ -131,6 +134,12 @@ def call(self, graph: List, **kwargs) -> tf.Tensor:
         """
         field = graph[getattr(Index, self.field.upper())]
         n_field = graph[getattr(Index, f"n_{self.field}".upper())]
+        if self.output_latent_feats is True:
+            return field, self.method_func(
+                field,
+                get_segment_indices_from_n(n_field),
+                num_segements=tf.shape(n_field)[0],
+            )
         return self.method_func(  # type: ignore
             field,
             get_segment_indices_from_n(n_field),

m3gnet/layers/_three_body.py

@@ -1,6 +1,7 @@
 """
 Three body basis expansion
 """
+
 from typing import List
 
 import tensorflow as tf

m3gnet/layers/_two_body.py

@@ -1,6 +1,7 @@
 """
 Calculate distance from atom positions and indices
 """
+
 from typing import List
 
 import tensorflow as tf

m3gnet/models/__init__.py

@@ -1,6 +1,7 @@
 """
 Graph pretrained
 """
+
 from ._base import BasePotential, GraphModelMixin, Potential
 from ._dynamics import M3GNetCalculator, MolecularDynamics, Relaxer
 from ._m3gnet import M3GNet

m3gnet/models/_dynamics.py

@@ -25,6 +25,7 @@
 from pymatgen.core.structure import Molecule, Structure
 from pymatgen.io.ase import AseAtomsAdaptor
 
+from m3gnet.layers import AtomRef
 from ._base import Potential
 from ._m3gnet import M3GNet
 
@@ -47,19 +48,31 @@ class M3GNetCalculator(Calculator):
 
     implemented_properties = ["energy", "free_energy", "forces", "stress"]
 
-    def __init__(self, potential: Potential, compute_stress: bool = True, stress_weight: float = 1.0, **kwargs):
+    def __init__(
+        self,
+        potential: Potential,
+        compute_stress: bool = True,
+        stress_weight: float = 1.0,
+        state_attr=None,
+        element_refs=None,
+        **kwargs,
+    ):
         """
 
         Args:
             potential (Potential): m3gnet.models.Potential
             compute_stress (bool): whether to calculate the stress
             stress_weight (float): the stress weight.
+            state_attr (np.ndarray): global state attribute (e.g. fidelity of data)
+            element_refs (np.ndarray): elemental energy offsets.
             **kwargs:
         """
         super().__init__(**kwargs)
         self.potential = potential
         self.compute_stress = compute_stress
         self.stress_weight = stress_weight
+        self.state_attr = state_attr
+        self.element_refs = AtomRef(property_per_element=element_refs) if element_refs is not None else element_refs
 
     def calculate(
         self,
@@ -81,14 +94,27 @@ def calculate(
         system_changes = system_changes or all_changes
         super().calculate(atoms=atoms, properties=properties, system_changes=system_changes)
 
-        graph = self.potential.graph_converter(atoms)
+        graph = self.potential.graph_converter(atoms, self.state_attr)
         graph_list = graph.as_tf().as_list()
         results = self.potential.get_efs_tensor(graph_list, include_stresses=self.compute_stress)
+        offset = None
+        if self.element_refs is not None:
+            offset = self.element_refs(graph_list)
+
         self.results.update(
-            energy=results[0].numpy().ravel()[0],
-            free_energy=results[0].numpy().ravel()[0],
+            energy=(
+                results[0].numpy().ravel()[0] + offset.numpy().ravel()[0]
+                if self.element_refs is not None
+                else results[0].numpy().ravel()[0]
+            ),
+            free_energy=(
+                results[0].numpy().ravel()[0] + offset.numpy().ravel()[0]
+                if self.element_refs is not None
+                else results[0].numpy().ravel()[0]
+            ),
             forces=results[1].numpy(),
         )
+
         if self.compute_stress:
             self.results.update(stress=results[2].numpy()[0] * self.stress_weight)
 
@@ -104,17 +130,21 @@ def __init__(
         optimizer: Union[Optimizer, str] = "FIRE",
         relax_cell: bool = True,
         stress_weight: float = 0.01,
+        state_attr=None,
+        element_refs=None,
     ):
         """
 
         Args:
             potential (Optional[Union[Potential, str]]): a potential,
                 a str path to a saved model or a short name for saved model
                 that comes with M3GNet distribution
-            optimizer (str or ase Optimizer): the optimization algorithm.
+            optimizer (str or ase Optimizer): the optimization algorithm
                 Defaults to "FIRE"
             relax_cell (bool): whether to relax the lattice cell
             stress_weight (float): the stress weight for relaxation
+            state_attr (np.ndarray): global state attribute (e.g. fidelity of data)
+            element_refs (np.ndarray): elemental energy offsets.
         """
         if isinstance(potential, str):
             potential = Potential(M3GNet.load(potential))
@@ -129,7 +159,9 @@ def __init__(
             optimizer_obj = optimizer
 
         self.opt_class: Optimizer = optimizer_obj
-        self.calculator = M3GNetCalculator(potential=potential, stress_weight=stress_weight)
+        self.calculator = M3GNetCalculator(
+            potential=potential, stress_weight=stress_weight, state_attr=state_attr, element_refs=element_refs
+        )
         self.relax_cell = relax_cell
         self.potential = potential
         self.ase_adaptor = AseAtomsAdaptor()
@@ -152,7 +184,7 @@ def relax(
                 Here fmax is a sum of force and stress forces
             steps (int): max number of steps for relaxation
             traj_file (str): the trajectory file for saving
-            interval (int): the step interval for saving the trajectories
+            interval (int): the step interval for saving the trajectories.
             **kwargs:
         Returns:
         """
@@ -257,6 +289,9 @@ def __init__(
         logfile: Optional[str] = None,
         loginterval: int = 1,
         append_trajectory: bool = False,
+        stress_weight: float = 1 / 160.21766208,
+        state_attr=None,
+        element_refs=None,
     ):
         """
 
@@ -276,6 +311,9 @@ def __init__(
             logfile (str): open this file for recording MD outputs
             loginterval (int): write to log file every interval steps
             append_trajectory (bool): Whether to append to prev trajectory
+            stress_weight (float): unit conversion from GPa to eV/A^3
+            state_attr (np.ndarray): global state attribute (e.g. fidelity of data)
+            element_refs (np.ndarray): elemental energy offsets.
         """
 
         if isinstance(potential, str):
@@ -284,7 +322,11 @@ def __init__(
         if isinstance(atoms, (Structure, Molecule)):
             atoms = AseAtomsAdaptor().get_atoms(atoms)
         self.atoms = atoms
-        self.atoms.set_calculator(M3GNetCalculator(potential=potential))
+        self.atoms.set_calculator(
+            M3GNetCalculator(
+                potential=potential, stress_weight=stress_weight, state_attr=state_attr, element_refs=element_refs
+            )
+        )
 
         if taut is None:
             taut = 100 * timestep * units.fs