add r_cut column to MetricsTable and make graph_construction_radius required key of hyperparams if model_type in (GNN, UIP)

add eSEN checkpoint URLs found at https://huggingface.co/facebook/OMAT24#models-trained-on-omat-mptrj-and-salexandria

Author: janosh

.pre-commit-config.yaml

@@ -8,7 +8,7 @@ default_install_hook_types: [pre-commit, commit-msg]
 
 repos:
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.11.4
+    rev: v0.11.5
     hooks:
       - id: ruff
         args: [--fix]
@@ -74,7 +74,7 @@ repos:
           - "@stylistic/eslint-plugin"
 
   - repo: https://github.com/python-jsonschema/check-jsonschema
-    rev: 0.32.1
+    rev: 0.33.0
     hooks:
       - id: check-jsonschema
         files: ^models/.+/.+\.yml$
@@ -85,7 +85,7 @@ repos:
       - id: check-github-actions
 
   - repo: https://github.com/RobertCraigie/pyright-python
-    rev: v1.1.398
+    rev: v1.1.399
     hooks:
       - id: pyright
         args: [--level, error]

contributing.md

@@ -153,7 +153,9 @@ To submit a new model to this benchmark and add it to our leaderboard, please cr
       max_steps: 500
       ase_optimizer: FIRE
       optimizer: Adam
-      ... # additional hyperparameters describing training
+      graph_construction_radius: 6.0
+      max_neighbors: 50
+      ... # additional hyperparameters describing training and inference
 
    training_cost: # list any hardware used to train the model and for how long
      # hardware: { amount: float, hours: float, cost: float [USD] }

matbench_discovery/modeling-tasks.yml

@@ -39,7 +39,7 @@ phonons:
     higher_is_better: []
   subtasks:
     kappa_103:
-      label: κ<sub>SRME</sub> (103)
+      label: κ<sub>SRME</sub>
       description: Lattice thermal conductivity symmetric relative mean error for 103 PhononDB structures
 
 diatomics:

models/alignn/alignn.yml

@@ -55,6 +55,11 @@ n_estimators: 1
 training_set: [MP 2022]
 training_cost: missing
 
+hyperparams:
+  # taken from alignn/alignn-mp22/alignn-config.json
+  graph_construction_radius: 8.0 # `cutoff` in Å, based on original implementation
+  max_neighbors: 12
+
 metrics:
   phonons: not applicable # reason: ALIGNN does not predict forces
   geo_opt: not applicable # reason: ALIGNN does not predict forces

models/alignn_ff/alignn-ff.yml

@@ -61,7 +61,10 @@ n_estimators: 1
 model_params: 0 # paper doesn't mention and no longer able to load model checkpoint to count parameters
 trained_for_benchmark: true
 status: aborted
-# for hyperparams, see alignn-config.json
+
+hyperparams:
+  graph_construction_radius: 8.0 # Å, from Methodology section in arXiv:2209.05554
+  max_neighbors: 12
 
 training_set: [MPtrj]
 training_cost: missing

models/alphanet/alphanet-mptrj.yml

@@ -80,7 +80,8 @@ hyperparams:
   n_hidden_channels: 256
   n_radial_guassian_basis: 256
   n_heads: 24
-  radial_cutoff: 6.0 # Å
+  graph_construction_radius: 6.0 # Å, from table 6 in arXiv:2501.07155
+  max_neighbors: .inf # max_num_neighbors_threshold seems to be unused in https://github.com/zmyybc/AlphaNet/blob/243fe71cb9/alphanet/models/graph.py#L28
 
 requirements:
   torch: 2.5.1

models/cgcnn/cgcnn+p.yml

@@ -55,6 +55,8 @@ training_cost: missing
 
 hyperparams:
   Perturbations: 5
+  graph_construction_radius: 5.0 # Å, from https://github.com/CompRhys/aviary/blob/451f5739/aviary/cgcnn/data.py#L28
+  max_neighbors: .inf # CGCNN paper benchmarks 6.0 Å cutoff radius vs. 12 NNs graph construction
 
 notes:
   Description: |

models/cgcnn/cgcnn.yml

@@ -50,6 +50,10 @@ n_estimators: 10
 training_set: [MP 2022]
 training_cost: missing
 
+hyperparams:
+  graph_construction_radius: 5.0 # Å, from  https://github.com/CompRhys/aviary/blob/451f5739/aviary/cgcnn/data.py#L28
+  max_neighbors: .inf # CGCNN paper benchmarks on 6.0 Å against 12 NNs graph construction
+
 notes:
   Description: |
     Published in 2018, CGCNN was the first crystal graph convolutional neural network to directly learn 8 different DFT-computed material properties from a graph representing the atoms and bonds in a crystal.

models/chgnet/chgnet-0.3.0.yml

@@ -66,6 +66,9 @@ hyperparams:
   ase_optimizer: FIRE
   cell_filter: FrechetCellFilter
   max_steps: 500
+  graph_construction_radius: 5.0 # Å, from sec. B. Model design in arXiv:2302.14231
+  three_body_cutoff: 3.0 # Å
+  max_neighbors: .inf
 
 notes:
   Description: |

models/deepmd/dpa3-v1-mptrj.yml

@@ -83,6 +83,8 @@ hyperparams:
     training_steps: 2000000
   batch_size: 64 # 16 (gpus) * 4 (batch per gpu) = 64 (total batch size)
   epochs: 160 # round1 80 + round2 80
+  graph_construction_radius: 6.0 # Å
+  max_neighbors: 120 # from https://github.com/deepmodeling/deepmd-kit/discussions/4682#discussioncomment-12836651
 
 requirements:
   torch: 2.3.1
@@ -96,7 +98,7 @@ training_cost: missing
 
 notes:
   Description: |
-    DPA3 is an advanced interatomic potential leveraging the message passing architecture, implemented within the DeePMD-kit framework, available at GitHub(https://github.com/deepmodeling/deepmd-kit/tree/dpa3-alpha).
+    DPA3 is an advanced interatomic potential leveraging the message passing architecture, implemented within the DeePMD-kit framework, available on [GitHub](https://github.com/deepmodeling/deepmd-kit/tree/dpa3-alpha).
     Designed as a large atomic model (LAM), DPA3 is tailored to integrate and simultaneously train on datasets from various disciplines, encompassing diverse chemical and materials systems across different research domains.
     Its model design ensures exceptional fitting accuracy and robust generalization both within and beyond the training domain.
     Furthermore, DPA3 maintains energy conservation and respects the physical symmetries of the potential energy surface, making it a dependable tool for a wide range of scientific applications.