Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
53 changes: 28 additions & 25 deletions interatomic_potentials/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -6,30 +6,33 @@ Machine-learning interatomic potentials (MLIP) bridge the gap between quantum-le

## 2.Models Matrix

| **Supported Functions** | **[CHGNet](./configs/chgnet/README.md)** | **[MatterSim](./configs/mattersim//README.md)** |
| ----------------------------------- | ---------------------------------------- | ----------------------------------------------- |
| **Forward Prediction** | | |
|  Energy | ✅ | ✅ |
|  Force | ✅ | ✅ |
|  Stress | ✅ | ✅ |
|  Magmom | ✅ | - |
| **ML Capabilities · Training** | | |
|  Single-GPU | ✅ | ✅ |
|  Distributed Train | ✅ | ✅ |
|  Mixed Precision | - | - |
|  Fine-tuning | ✅ | ✅ |
|  Uncertainty / Active-Learning | - | - |
|  Dynamic→Static | - | - |
|  Compiler CINN | - | - |
| **ML Capabilities · Predict** | | |
|  Distillation / Pruning | - | - |
|  Standard inference | ✅ | ✅ |
|  Distributed inference | - | - |
|  Compiler CINN | - | - |
| **Molecular Dynamic Interface** | | |
|  ASE | ✅ | ✅ |
| **Dataset** | | |
|  MPtrj | ✅ | 🚧 |
| **ML2DDB🌟** | ✅ | - |
| **Supported Functions** | **[CHGNet](./configs/chgnet/README.md)** | **[MatterSim](./configs/mattersim//README.md)** | **[SchNet](./configs/schnet/README.md)** |
| ----------------------------------- | ---------------------------------------- | ----------------------------------------------- | ---------------------------------------- |
| **Forward Prediction** | | | |
|  Energy | ✅ | ✅ | ✅ |
|  Force | ✅ | ✅ | ✅ |
|  Stress | ✅ | ✅ | - |
|  Magmom | ✅ | - | - |
| **ML Capabilities · Training** | | | |
|  Single-GPU | ✅ | ✅ | ✅ |
|  Distributed Train | ✅ | ✅ | ✅ |
|  Mixed Precision | - | - | - |
|  Fine-tuning | ✅ | ✅ | ✅ |
|  Uncertainty / Active-Learning | - | - | - |
|  Dynamic→Static | - | - | - |
|  Compiler CINN | - | - | - |
| **ML Capabilities · Predict** | | | |
|  Distillation / Pruning | - | - | - |
|  Standard inference | ✅ | ✅ | ✅ |
|  Distributed inference | - | - | - |
|  Compiler CINN | - | - | - |
| **Molecular Dynamic Interface** | | | |
|  ASE | ✅ | ✅ | - |
| **Dataset** | | | |
|  MPtrj | ✅ | 🚧 | - |
|  QM9 | - | - | ✅ |
|  MD17 | - | - | ✅ |
|  ISO17 | - | - | ✅ |
| **ML2DDB🌟** | ✅ | - | - |

**Notice**:🌟 represent originate research work published from paddlematerials toolkit
187 changes: 187 additions & 0 deletions interatomic_potentials/configs/schnet/README.md
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

辛苦补充完整results表格

Original file line number Diff line number Diff line change
@@ -0,0 +1,187 @@
# SchNet

[SchNet: A continuous-filter convolutional neural network for modeling quantum interactions](https://arxiv.org/abs/1706.08566)

## Abstract

SchNet is an end-to-end continuous-filter convolutional neural network for atomistic systems. It models quantum interactions directly from atomic numbers and coordinates, using distance-based filters with smooth cutoffs and atom-wise readout. In PaddleMaterials, this implementation targets paper-style small-molecule regression benchmarks and keeps training/inference interfaces consistent with the `interatomic_potentials` suite.

## Datasets

This SchNet case currently covers:

- QM9 (`U0`) with paper-aligned split setup
- MD17 (ethanol example, energy + force supervision)
- ISO17 (reference to test_other generalization)

Dataset sources:

- **QM9 (small molecules)**:
- Original source: https://figshare.com/ndownloader/files/3195389
- PaddleMaterials packaged mirror (used by SchNet config):
`https://paddle-org.bj.bcebos.com/paddlematerials/datasets/qm9/qm9.tar.gz`
- PaddleMaterials raw mirror (fallback / non-packaged path):
`https://paddle-org.bj.bcebos.com/paddlematerials/datasets/qm9/dsgdb9nsd.xyz.tar.bz2`
- **MD17 / rMD17 (small molecules, energy + force)**:
- Original source: https://www.quantum-machine.org/datasets/
- PaddleMaterials package mirror:
`https://paddle-org.bj.bcebos.com/paddlematerials/datasets/MD17/md17.tar.gz`
- Recommended format in this repo: `*.npz` with keys `R/Z/E/F`
- **ISO17 (isomer generalization benchmark)**:
- Original source: https://www.quantum-machine.org/datasets/
- PaddleMaterials package mirror:
`https://paddle-org.bj.bcebos.com/paddlematerials/datasets/ISO17/iso17.tar.gz`
- Recommended format in this repo: `iso17.npz` with keys `R/Z/E` and optional `isomer_ids`

## Models

The Paddle implementation follows the SchNet interaction design:

- Gaussian RBF expansion for inter-atomic distances
- Cosine cutoff for locality
- Continuous-filter interaction blocks with residual updates
- Atom-wise readout for molecular properties

Config files:

- `schnet_qm9_lumo.yaml`: paper-aligned QM9 `U0` case (filename kept for compatibility)
- `schnet_md17_ethanol.yaml`: MD17 ethanol energy-force joint training
- `schnet_iso17.yaml`: ISO17 reference->test_other setting

Training hyperparameters in these SchNet configs are aligned to
SchNet-master/paper style (instead of current schnetpack defaults):

- Optimizer: `Adam`
- LR schedule: exponential decay (`lr=1e-3`, decay steps `100000`, gamma `0.96`)
- Global-step stop: `max_iter=5000000`
- Step-based validation/save: `eval_interval_steps=5000`, `save_interval_steps=50000`
- QM9 train batch: `32`, val/test batch: `100`
- QM9 split file: `split_qm9_110k_1k_seed42.npz` (`num_train=110000`, `num_val=1000`)

## Results

<table>
<head>
<tr>
<th nowrap="nowrap">Model Name</th>
<th nowrap="nowrap">Dataset</th>
<th nowrap="nowrap">Task</th>
<th nowrap="nowrap">Metric</th>
<th nowrap="nowrap">GPUs</th>
<th nowrap="nowrap">Training time</th>
<th nowrap="nowrap">Config</th>
<th nowrap="nowrap">Checkpoint | Log</th>
</tr>
</head>
<body>
<tr>
<td nowrap="nowrap">schnet_qm9_u0</td>
<td nowrap="nowrap">QM9</td>
<td nowrap="nowrap">U0 regression</td>
<td nowrap="nowrap">MAE: to_be_filled</td>
<td nowrap="nowrap">~</td>
<td nowrap="nowrap">~</td>
<td nowrap="nowrap"><a href="schnet_qm9_lumo.yaml">schnet_qm9_lumo.yaml</a></td>
<td nowrap="nowrap"><a href="to_be_filled">checkpoint | log</a></td>
</tr>
<tr>
<td nowrap="nowrap">schnet_md17_ethanol</td>
<td nowrap="nowrap">MD17</td>
<td nowrap="nowrap">Energy + Force</td>
<td nowrap="nowrap">MAE: to_be_filled</td>
<td nowrap="nowrap">~</td>
<td nowrap="nowrap">~</td>
<td nowrap="nowrap"><a href="schnet_md17_ethanol.yaml">schnet_md17_ethanol.yaml</a></td>
<td nowrap="nowrap"><a href="to_be_filled">checkpoint | log</a></td>
</tr>
<tr>
<td nowrap="nowrap">schnet_iso17</td>
<td nowrap="nowrap">ISO17</td>
<td nowrap="nowrap">Energy regression</td>
<td nowrap="nowrap">MAE: to_be_filled</td>
<td nowrap="nowrap">~</td>
<td nowrap="nowrap">~</td>
<td nowrap="nowrap"><a href="schnet_iso17.yaml">schnet_iso17.yaml</a></td>
<td nowrap="nowrap"><a href="to_be_filled">checkpoint | log</a></td>
</tr>
</body>
</table>

## Alignment Checklist

Use the following acceptance criteria during migration:

- Single-card forward alignment: logits diff around `1e-4` (generative models `1e-6`)
- Backward alignment: after at least 2 epochs, train loss trend is consistent
- Supervised task metric error: within `1%`

Quick local checks:

```bash
python test/run_schnet_torch_alignment.py
python test/run_schnet_data_metric_alignment.py
```

Dataset preparation from original/torch sources:

```bash
python test/prepare_schnet_paper_datasets.py --data_root ./data --tasks all
```

## Resource Links (Baidu)

Per PaddleMaterials delivery requirement, fill cloud links here after PR/WeChat handoff:

- Dataset package links:
- QM9: `https://paddle-org.bj.bcebos.com/paddlematerials/datasets/qm9/qm9.tar.gz`
- MD17: `https://paddle-org.bj.bcebos.com/paddlematerials/datasets/MD17/md17.tar.gz`
- ISO17: `https://paddle-org.bj.bcebos.com/paddlematerials/datasets/ISO17/iso17.tar.gz`
- Pretrained model link (Baidu): `to_be_filled`
- Training/inference raw log link (Baidu): `to_be_filled`

### Training

```bash
# multi-gpu training (QM9)
python -m paddle.distributed.launch --gpus="0,1,2,3" interatomic_potentials/train.py -c interatomic_potentials/configs/schnet/schnet_qm9_lumo.yaml

# single-gpu training (QM9)
python interatomic_potentials/train.py -c interatomic_potentials/configs/schnet/schnet_qm9_lumo.yaml

# MD17 (ethanol)
python interatomic_potentials/train.py -c interatomic_potentials/configs/schnet/schnet_md17_ethanol.yaml

# ISO17
python interatomic_potentials/train.py -c interatomic_potentials/configs/schnet/schnet_iso17.yaml
```

### Validation

```bash
python interatomic_potentials/train.py -c interatomic_potentials/configs/schnet/schnet_qm9_lumo.yaml Global.do_eval=True Global.do_train=False Global.do_test=False Trainer.pretrained_model_path='your checkpoint path(*.pdparams)'
```

### Testing

```bash
python interatomic_potentials/train.py -c interatomic_potentials/configs/schnet/schnet_qm9_lumo.yaml Global.do_test=True Global.do_train=False Global.do_eval=False Trainer.pretrained_model_path='your checkpoint path(*.pdparams)'
```

### Prediction

```bash
# Mode 1: load from local config/checkpoint
python interatomic_potentials/predict.py --config_path='interatomic_potentials/configs/schnet/schnet_qm9_lumo.yaml' --checkpoint_path='your checkpoint path(*.pdparams)' --cif_file_path='./interatomic_potentials/example_data/cifs/'
```

## Citation

```
@inproceedings{schutt2017schnet,
title={SchNet: A continuous-filter convolutional neural network for modeling quantum interactions},
author={Sch{\"u}tt, Kristof T and Kindermans, Pieter-Jan and Sauceda, Felix A and Chmiela, Stefan and Tkatchenko, Alexandre and M{\"u}ller, Klaus-Robert},
booktitle={Advances in Neural Information Processing Systems},
volume={30},
year={2017}
}
```
141 changes: 141 additions & 0 deletions interatomic_potentials/configs/schnet/schnet_iso17.yaml
Original file line number Diff line number Diff line change
@@ -0,0 +1,141 @@
Global:
do_train: True
do_eval: True
do_test: True

label_names: ["energy"]
graph_converter:
__class_name__: FindPointsInSpheres
__init_params__:
# Paper-priority setting.
cutoff: 5.0
pbc: [0, 0, 0]
num_cpus: 4


Trainer:
# Paper + SchNet-master aligned optimizer style.
max_epochs: 100000
# SchNet-master style stop/eval/save by global step.
max_iter: 5000000
eval_interval_steps: 5000
save_interval_steps: 50000
seed: 42
output_dir: ./output/schnet_iso17_paper
save_freq: 100
log_freq: 20
start_eval_epoch: 1
eval_freq: 1
pretrained_model_path: null
pretrained_weight_name: null
resume_from_checkpoint: null
use_amp: False
amp_level: "O1"
eval_with_no_grad: True
gradient_accumulation_steps: 1
best_metric_indicator: "eval_metric"
name_for_best_metric: "energy"
greater_is_better: False
compute_metric_during_train: True
metric_strategy_during_eval: "epoch"
use_visualdl: False
use_wandb: False
use_tensorboard: False


Model:
__class_name__: SchNet
__init_params__:
# Paper-priority setting.
n_atom_basis: 128
n_interactions: 6
n_filters: 128
cutoff: 5.0
n_rbf: 20
max_z: 100
property_name: "energy"
readout: "sum"
data_mean: 0.0
data_std: 1.0
loss_type: "mse_loss"


Optimizer:
__class_name__: Adam
__init_params__:
lr:
__class_name__: ExponentialDecay
__init_params__:
learning_rate: 1e-3
gamma: 0.96
decay_steps: 100000
by_epoch: False


Metric:
energy:
__class_name__: paddle.nn.L1Loss
__init_params__: {}


Dataset:
# Dataset package mirror (uploaded):
# https://paddle-org.bj.bcebos.com/paddlematerials/datasets/ISO17/iso17.tar.gz
train:
dataset:
__class_name__: ISO17Dataset
__init_params__:
path: "./data/iso17"
url: "https://paddle-org.bj.bcebos.com/paddlematerials/datasets/ISO17/iso17.tar.gz"
subset: "train"
split_file: "split_iso17_ref90_testother_seed42.npz"
property_names: ${Global.label_names}
build_graph_cfg: ${Global.graph_converter}
seed: 42
sampler:
__class_name__: BatchSampler
__init_params__:
shuffle: True
drop_last: True
batch_size: 32

val:
dataset:
__class_name__: ISO17Dataset
__init_params__:
path: "./data/iso17"
url: "https://paddle-org.bj.bcebos.com/paddlematerials/datasets/ISO17/iso17.tar.gz"
subset: "val"
split_file: "split_iso17_ref90_testother_seed42.npz"
property_names: ${Global.label_names}
build_graph_cfg: ${Global.graph_converter}
seed: 42
sampler:
__class_name__: BatchSampler
__init_params__:
shuffle: False
drop_last: False
batch_size: 100

test:
dataset:
__class_name__: ISO17Dataset
__init_params__:
path: "./data/iso17"
url: "https://paddle-org.bj.bcebos.com/paddlematerials/datasets/ISO17/iso17.tar.gz"
subset: "test"
split_file: "split_iso17_ref90_testother_seed42.npz"
property_names: ${Global.label_names}
build_graph_cfg: ${Global.graph_converter}
seed: 42
sampler:
__class_name__: BatchSampler
__init_params__:
shuffle: False
drop_last: False
batch_size: 100


Predict:
graph_converter: ${Global.graph_converter}
eval_with_no_grad: True
Loading