Merge branch 'PaddlePaddle:develop' into dev_model

Author: Nolan-149

README.md

@@ -108,7 +108,7 @@ PaddleScience 是一个基于深度学习框架 PaddlePaddle 开发的科学计
 | 晶体材料属性预测 | [CGCNN](https://paddlescience-docs.readthedocs.io/zh-cn/latest/zh/examples/cgcnn/) | 数据驱动 | GNN | 监督学习 | [MP](https://next-gen.materialsproject.org/) / [Perovskite](https://cmr.fysik.dtu.dk/cubic_perovskites/cubic_perovskites.html) / [C2DB](https://cmr.fysik.dtu.dk/c2db/c2db.html) / [test](https://paddle-org.bj.bcebos.com/paddlescience%2Fdatasets%2Fcgcnn%2Fcgcnn-test.zip) | [Paper](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.120.145301) |
 | 分子生成 | [MoFlow](https://paddlescience-docs.readthedocs.io/zh-cn/latest/zh/examples/moflow/) | 数据驱动 | Flow Model | 监督学习 | [qm9/ zink250k](https://aistudio.baidu.com/datasetdetail/282687) | [Paper](https://arxiv.org/abs/2006.10137v1) |
 | 分子属性预测 | [IFM](https://paddlescience-docs.readthedocs.io/zh-cn/latest/zh/examples/ifm/) | 数据驱动 | MLP | 监督学习 | [tox21/sider/hiv/bace/bbbp](https://paddlescience-docs.readthedocs.io/zh-cn/latest/zh/examples/ifm/#:~:text=molecules%20%E6%95%B0%E6%8D%AE%E9%9B%86-,dataset.zip,-%EF%BC%8C%E6%88%96Google%20Drive) | [Paper](https://openreview.net/pdf?id=NLFqlDeuzt) |
-
+| 二维材料生成与数据库 | [ML2DDB](./en/examples/ml2ddb.md) | 数据驱动 | GNN/Diffusion | 监督学习 | Coming Soon | [Paper](https://arxiv.org/pdf/2507.00584) |
 
 <br>
 <p align="center"><b>地球科学(AI for Earth Science)</b></p>
@@ -166,7 +166,7 @@ PaddleScience 是一个基于深度学习框架 PaddlePaddle 开发的科学计
 ### 安装 PaddlePaddle
 
 <!-- --8<-- [start:paddle_install] -->
-请根据您的运行环境，访问 [PaddlePaddle](https://www.paddlepaddle.org.cn/install/quick?docurl=/documentation/docs/zh/develop/install/pip/linux-pip.html) 官网，安装 <font color="red"><b>3.0 或 develop</b></font> 版的 PaddlePaddle。
+请根据您的运行环境，访问 [PaddlePaddle](https://www.paddlepaddle.org.cn/install/quick?docurl=/documentation/docs/zh/develop/install/pip/linux-pip.html) 官网，建议安装 PaddlePaddle <font color="red"><b>3.0 以上稳定版，或最新的 develop 开发版</b></font>。
 
 安装完毕之后，运行以下命令，验证 Paddle 是否安装成功。
 

deploy/python_infer/base.py

@@ -39,7 +39,7 @@ class Predictor:
         pdmodel_path (Optional[str]): Path to the PaddlePaddle model file. Defaults to None.
         pdiparams_path (Optional[str]): Path to the PaddlePaddle model parameters file. Defaults to None.
         device (Literal["cpu", "gpu", "npu", "xpu", "sdaa"], optional): Device to use for inference. Defaults to "cpu".
-        engine (Literal["native", "tensorrt", "onnx", "mkldnn"], optional): Inference engine to use. Defaults to "native".
+        engine (Literal["native", "tensorrt", "onnx", "onednn"], optional): Inference engine to use. Defaults to "native".
         precision (Literal["fp32", "fp16", "int8"], optional): Precision to use for inference. Defaults to "fp32".
         onnx_path (Optional[str], optional): Path to the ONNX model file. Defaults to None.
         ir_optim (bool, optional): Whether to use IR optimization. Defaults to True.
@@ -55,7 +55,7 @@ def __init__(
         pdiparams_path: Optional[str] = None,
         *,
         device: Literal["cpu", "gpu", "npu", "xpu", "sdaa"] = "cpu",
-        engine: Literal["native", "tensorrt", "onnx", "mkldnn"] = "native",
+        engine: Literal["native", "tensorrt", "onnx", "onednn"] = "native",
         precision: Literal["fp32", "fp16", "int8"] = "fp32",
         onnx_path: Optional[str] = None,
         ir_optim: bool = True,
@@ -157,11 +157,11 @@ def _create_paddle_predictor(
             config.enable_xpu(10 * 1024 * 1024)
         else:
             config.disable_gpu()
-            if self.engine == "mkldnn":
+            if self.engine == "onednn":
                 # 'set_mkldnn_cache_capatity' is not available on macOS
                 if platform.system() != "Darwin":
                     ...
-                    # cache 10 different shapes for mkldnn to avoid memory leak
+                    # cache 10 different shapes for onednn to avoid memory leak
                     # config.set_mkldnn_cache_capacity(10)
                 config.enable_mkldnn()
 
@@ -170,6 +170,11 @@ def _create_paddle_predictor(
 
                 config.set_cpu_math_library_num_threads(self.num_cpu_threads)
 
+            elif self.engine == "mkldnn":
+                raise ValueError(
+                    "The 'mkldnn' engine is deprecated. Please use 'onednn' instead."
+                )
+
         # enable memory optim
         config.enable_memory_optim()
         # config.disable_glog_info()
@@ -221,9 +226,9 @@ def _check_device(self, device: str):
             )
 
     def _check_engine(self, engine: str):
-        if engine not in ["native", "tensorrt", "onnx", "mkldnn"]:
+        if engine not in ["native", "tensorrt", "onnx", "onednn"]:
             raise ValueError(
-                "Inference only supports 'native', 'tensorrt', 'onnx' and 'mkldnn' "
+                "Inference only supports 'native', 'tensorrt', 'onnx' and 'onednn' "
                 f"engines, but got {engine}."
             )
 

docs/en/examples/ml2ddb.md

@@ -0,0 +1,59 @@
+# ML2DDB
+
+[Monolayer Two-dimensional Materials Database (ML2DDB) and Applications](https://arxiv.org/pdf/2507.00584)
+
+Zhongwei Liu<sup>a, b, #</sup>,
+Zhimin Zhang<sup>c, #</sup>,
+Xuwei Liu<sup>c, #</sup>,
+Mingjia Yao<sup>b</sup>,
+Xin He<sup>a</sup>,
+Yuanhui Sun<sup>b, *</sup>,
+Xin Chen<sup>b, *</sup>,
+Lijun Zhang<sup>a, b, *</sup>
+
+<sup>a</sup>
+State Key Laboratory of Integrated Optoelectronics, Key Laboratory of Automobile Materials of MOE and College of Materials Science and Engineering, Jilin University, Changchun 130012, China
+
+<sup>b</sup> Suzhou Laboratory, Suzhou, 215123, China
+
+<sup>c</sup> Baidu Inc., Beijing, P.R. China.
+
+<sup>#</sup> These authors contributed equally to this work.
+
+E-mail: [email protected]; [email protected]; [email protected]
+
+## Abstract
+
+The discovery of two-dimensional (2D) materials with tailored properties is critical to meet the increasing demands of high-performance applications across flexible electronics, optoelectronics, catalysis, and energy storage. However, current 2D material databases are constrained by limited scale and compositional diversity. In this study, we introduce a scalable active learning workflow that integrates deep neural networks with density functional theory (DFT) calculations to efficiently explore a vast set of candidate structures. These structures are generated through physics-informed elemental substitution strategies, enabling broad and systematic discovery of stable 2D materials. Through six iterative screening cycles, we established the creation of the Monolayer 2D Materials Database (ML2DDB), which contains 242,546 DFT-validated stable structures—an order-of-magnitude increase over the largest known 2D materials databases. In particular, the number of ternary and quaternary compounds showed the most significant increase. Combining this database with a generative diffusion model, we demonstrated effective structure generation under specified chemistry and symmetry constraints. This work accomplished an organically interconnected loop of 2D material data expansion and application, which provides a new paradigm for the discovery of new materials.
+
+![ML2DDB](https://paddle-org.bj.bcebos.com/paddlescience/docs/ML2DDB/ml2ddb.png)
+
+## Dataset of 2D materials
+
+We developed ML2DDB, a large-scale 2D material database containing >242k DFT-validated monolayer structures (𝐸<sub>hull</sub><sup>𝐷𝐹𝑇</sup> <50 meV/atom), representing a 10× increase over existing datasets. Key features:
+
+- Broad elemental coverage: 81 elements across the periodic table (excluding radioactive/noble gases).
+- Enhanced diversity: Significantly more compounds with 3–4 distinct elements compared to prior work.
+- Structural richness: Diverse prototypes and cation-anion combinations.
+- Extended resource: >1M candidate structures (𝐸<sub>hull</sub><sup>MLIP</sup> <200 meV/atom) for future studies.
+
+![dataset](https://paddle-org.bj.bcebos.com/paddlescience/docs/ML2DDB/ml2ddb_dataset.png)
+
+## Diffusion model generation of S.U.N. materials
+
+The capability to generate S.U.N. (stable, unique, new) 2D materials are prerequisites for diffusion models. We considered a generated structure as stable with 𝐸<sub>hull</sub><sup>𝐷𝐹𝑇</sup> < 100 meV/atom with respect to ML2DDB. The unique is specified whether a generated structure matches any other structure generated in the same batch or not, and the new is whether it is identical to any of the structures in ML2DDB. As shown in Figure 5b, we performed DFT structure optimization on 1024 structures to evaluate the stable attribute. The results show that 74.8% of them are considered stable (𝐸<sub>hull</sub><sup>𝐷𝐹𝑇</sup> < 100 meV/atom), which is comparable to the success rate of 3D stable structure generation of MatterGen. When the constraint is set to 𝐸<sub>hull</sub><sup>𝐷𝐹𝑇</sup> < 0 meV/atom, our method achieved a success rate of 59.6%, which is significantly higher than that of  MatterGen (~13%). In addition, the Root-mean-square displacement (RMSD) of the generated structure is lower than 0.26 Å compared to the DFT relaxation structure, which is still less than the radius of the hydrogen atom (0.53 Å). For the generation of unique structures, the success rate accounts for 100% when generating one thousand structures. The rate only decreases 4.4% when generating ten thousand structures. For the generation of new structures, the rate decreases from 100% to 73.5% when the generated structures grow from one thousand to two thousand. This indicates that our model has a relatively excellent ability to generate completely new stable structures.
+
+![dataset](https://paddle-org.bj.bcebos.com/paddlescience/docs/ML2DDB/gen_2d.png)
+
+## Conclusion
+
+This study establishes a novel framework integrating active learning workflows with conditional diffusion-based structural generation, achieving unprecedented expansion of 2D materials databases. Key contributions include:
+
+1. **Dataset Advancement**
+   - Created ML2DDB containing >242,546 thermodynamically stable 2D materials (E_hull^DFT <50 meV/atom), exceeding existing databases by ≥10x
+   - Achieved 1100% and 960% growth in ternary/quaternary compounds respectively
+   - Generated >1 million candidate structures (𝐸<sub>hull</sub><sup>MLIP</sup> <200 meV/atom)
+2. **Methodological Innovation**
+   - Developed MLIP model with 92.36% accuracy in stability classification
+   - Enabled phase diagram generation and space group-specific design through diffusion model integration
+   - Demonstrated applicability to nonlinear optical and ferroelectric materials discovery

docs/index.md

@@ -141,6 +141,7 @@
 |-----|---------|-----|---------|----|---------|---------|
 | 材料设计 | [散射板设计(反问题)](./zh/examples/hpinns.md) | 机理驱动 | Transformer | 无监督学习 | [Train Data](https://paddle-org.bj.bcebos.com/paddlescience/datasets/hPINNs/hpinns_holo_train.mat)<br>[Eval Data](https://paddle-org.bj.bcebos.com/paddlescience/datasets/hPINNs/hpinns_holo_valid.mat) | [Paper](https://arxiv.org/pdf/2102.04626.pdf) |
 | 晶体材料属性预测 | [CGCNN](./zh/examples/cgcnn.md) | 数据驱动 | GNN | 监督学习 | [MP](https://next-gen.materialsproject.org/) / [Perovskite](https://cmr.fysik.dtu.dk/cubic_perovskites/cubic_perovskites.html) / [C2DB](https://cmr.fysik.dtu.dk/c2db/c2db.html) / [test](https://paddle-org.bj.bcebos.com/paddlescience%2Fdatasets%2Fcgcnn%2Fcgcnn-test.zip) | [Paper](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.120.145301) |
+| 二维材料生成与数据库 | [ML2DDB](./en/examples/ml2ddb.md) | 数据驱动 | GNN/Diffusion | 监督学习 | Coming Soon | [Paper](https://arxiv.org/pdf/2507.00584) |
 
 <br>
 <p align="center"><b>地球科学(AI for Earth Science)</b></p>

docs/zh/api/utils/initializer.md

@@ -16,5 +16,7 @@
         - kaiming_normal_
         - linear_init_
         - conv_init_
+        - glorot_normal_
+        - lecun_normal_
       show_root_heading: True
       heading_level: 3

docs/zh/user_guide.md

@@ -464,7 +464,7 @@ ppsci MESSAGE: Visualization result is saved to: ./aneurysm_pred.vtu
 
 PaddleScience 提供了多种推理配置组合，可通过命令行进行组合，目前支持的推理配置如下：
 
-|  | Native | ONNX | TensorRT | macaRT | MKLDNN |
+|  | Native | ONNX | TensorRT | macaRT | oneDNN |
 | :--- | :--- | :--- | :--- | :--- | :--- |
 | Intel(CPU) | ✅ | ✅ | / | / | ✅ |
 | NVIDIA | ✅ | ✅ | ✅ | / | / |
@@ -576,31 +576,31 @@ PaddleScience 提供了多种推理配置组合，可通过命令行进行组合
         INFER.engine=onnx
     ```
 
-=== "使用 MKLDNN 推理"
+=== "使用 oneDNN 推理"
 
-    MKLDNN 是英特尔推出的高性能推理引擎，适用于 CPU 推理加速，PaddleScience 支持了 MKLDNN 推理功能。
+    oneDNN 是英特尔推出的高性能推理引擎，适用于 CPU 推理加速，PaddleScience 支持了 oneDNN 推理功能。
 
     运行以下命令进行推理：
 
     ``` sh
     python aneurysm.py mode=infer \
         INFER.device=cpu \
-        INFER.engine=mkldnn
+        INFER.engine=onednn
     ```
 
 !!! info "完整推理配置参数"
 
     | 参数 | 默认值 | 说明 |
     | :--- | :--- | :--- |
     | `INFER.device` | `cpu` | 推理设备，目前支持 `cpu` 和 `gpu` |
-    | `INFER.engine` | `native` | 推理引擎，目前支持 `native`, `tensorrt`, `onnx` 和 `mkldnn` |
+    | `INFER.engine` | `native` | 推理引擎，目前支持 `native`, `tensorrt`, `onnx` 和 `onednn` |
     | `INFER.precision` | `fp32` | 推理精度，目前支持 `fp32`, `fp16` |
     | `INFER.ir_optim` | `True` | 是否启用 IR 优化 |
     | `INFER.min_subgraph_size` | `30` | TensorRT 中最小子图 size，当子图的 size 大于该值时，才会尝试对该子图使用 TensorRT 计算 |
     | `INFER.gpu_mem` | `2000` | 初始显存大小 |
     | `INFER.gpu_id` | `0` | GPU 逻辑设备号 |
     | `INFER.max_batch_size` | `1024` | 推理时的最大 batch_size |
-    | `INFER.num_cpu_threads` | `10` | MKLDNN 和 ONNX 在 CPU 推理时的线程数 |
+    | `INFER.num_cpu_threads` | `10` | oneDNN 和 ONNX 在 CPU 推理时的线程数 |
     | `INFER.batch_size` | `256` | 推理时的 batch_size |
 
 ### 1.4 断点继续训练

examples/ML2DDB/readme.md

@@ -0,0 +1,59 @@
+# ML2DDB
+
+[Monolayer Two-dimensional Materials Database (ML2DDB) and Applications](https://arxiv.org/pdf/2507.00584)
+
+Zhongwei Liu<sup>a, b, #</sup>,
+Zhimin Zhang<sup>c, #</sup>,
+Xuwei Liu<sup>c, #</sup>,
+Mingjia Yao<sup>b</sup>,
+Xin He<sup>a</sup>,
+Yuanhui Sun<sup>b, *</sup>,
+Xin Chen<sup>b, *</sup>,
+Lijun Zhang<sup>a, b, *</sup>
+
+<sup>a</sup>
+State Key Laboratory of Integrated Optoelectronics, Key Laboratory of Automobile Materials of MOE and College of Materials Science and Engineering, Jilin University, Changchun 130012, China
+
+<sup>b</sup> Suzhou Laboratory, Suzhou, 215123, China
+
+<sup>c</sup> Baidu Inc., Beijing, P.R. China.
+
+<sup>#</sup> These authors contributed equally to this work.
+
+E-mail: [email protected]; [email protected]; [email protected]
+
+## Abstract
+
+The discovery of two-dimensional (2D) materials with tailored properties is critical to meet the increasing demands of high-performance applications across flexible electronics, optoelectronics, catalysis, and energy storage. However, current 2D material databases are constrained by limited scale and compositional diversity. In this study, we introduce a scalable active learning workflow that integrates deep neural networks with density functional theory (DFT) calculations to efficiently explore a vast set of candidate structures. These structures are generated through physics-informed elemental substitution strategies, enabling broad and systematic discovery of stable 2D materials. Through six iterative screening cycles, we established the creation of the Monolayer 2D Materials Database (ML2DDB), which contains 242,546 DFT-validated stable structures—an order-of-magnitude increase over the largest known 2D materials databases. In particular, the number of ternary and quaternary compounds showed the most significant increase. Combining this database with a generative diffusion model, we demonstrated effective structure generation under specified chemistry and symmetry constraints. This work accomplished an organically interconnected loop of 2D material data expansion and application, which provides a new paradigm for the discovery of new materials.
+
+![ML2DDB](https://paddle-org.bj.bcebos.com/paddlescience/docs/ML2DDB/ml2ddb.png)
+
+## Dataset of 2D materials
+
+We developed ML2DDB, a large-scale 2D material database containing >242k DFT-validated monolayer structures (𝐸<sub>hull</sub><sup>𝐷𝐹𝑇</sup> <50 meV/atom), representing a 10× increase over existing datasets. Key features:
+
+- Broad elemental coverage: 81 elements across the periodic table (excluding radioactive/noble gases).
+- Enhanced diversity: Significantly more compounds with 3–4 distinct elements compared to prior work.
+- Structural richness: Diverse prototypes and cation-anion combinations.
+- Extended resource: >1M candidate structures (𝐸<sub>hull</sub><sup>MLIP</sup> <200 meV/atom) for future studies.
+
+![dataset](https://paddle-org.bj.bcebos.com/paddlescience/docs/ML2DDB/ml2ddb_dataset.png)
+
+## Diffusion model generation of S.U.N. materials
+
+The capability to generate S.U.N. (stable, unique, new) 2D materials are prerequisites for diffusion models. We considered a generated structure as stable with 𝐸<sub>hull</sub><sup>𝐷𝐹𝑇</sup> < 100 meV/atom with respect to ML2DDB. The unique is specified whether a generated structure matches any other structure generated in the same batch or not, and the new is whether it is identical to any of the structures in ML2DDB. As shown in Figure 5b, we performed DFT structure optimization on 1024 structures to evaluate the stable attribute. The results show that 74.8% of them are considered stable (𝐸<sub>hull</sub><sup>𝐷𝐹𝑇</sup> < 100 meV/atom), which is comparable to the success rate of 3D stable structure generation of MatterGen. When the constraint is set to 𝐸<sub>hull</sub><sup>𝐷𝐹𝑇</sup> < 0 meV/atom, our method achieved a success rate of 59.6%, which is significantly higher than that of  MatterGen (~13%). In addition, the Root-mean-square displacement (RMSD) of the generated structure is lower than 0.26 Å compared to the DFT relaxation structure, which is still less than the radius of the hydrogen atom (0.53 Å). For the generation of unique structures, the success rate accounts for 100% when generating one thousand structures. The rate only decreases 4.4% when generating ten thousand structures. For the generation of new structures, the rate decreases from 100% to 73.5% when the generated structures grow from one thousand to two thousand. This indicates that our model has a relatively excellent ability to generate completely new stable structures.
+
+![dataset](https://paddle-org.bj.bcebos.com/paddlescience/docs/ML2DDB/gen_2d.png)
+
+## Conclusion
+
+This study establishes a novel framework integrating active learning workflows with conditional diffusion-based structural generation, achieving unprecedented expansion of 2D materials databases. Key contributions include:
+
+1. **Dataset Advancement**
+   - Created ML2DDB containing >242,546 thermodynamically stable 2D materials (E_hull^DFT <50 meV/atom), exceeding existing databases by ≥10x
+   - Achieved 1100% and 960% growth in ternary/quaternary compounds respectively
+   - Generated >1 million candidate structures (𝐸<sub>hull</sub><sup>MLIP</sup> <200 meV/atom)
+2. **Methodological Innovation**
+   - Developed MLIP model with 92.36% accuracy in stability classification
+   - Enabled phase diagram generation and space group-specific design through diffusion model integration
+   - Demonstrated applicability to nonlinear optical and ferroelectric materials discovery