【Hackathon 10th Spring No.7】wD-MPNN模型复现

[r-cloudforge]

通过迁移 https://github.com/coleygroup/polymer-chemprop 项目实现 wD-MPNN (weighted Directed Message Passing Neural Network) 的PaddlePaddle版本复现。

主要改动

1. 新增 ppmat/models/wd_mpnn/ 模型目录 — 完整的wD-MPNN实现，包括加权有向消息传递编码器和聚合物性质预测
2. 新增 property_prediction/configs/wd_mpnn/ 训练配置
3. 新增 test/wd_mpnn/ 测试代码 — 8个单元测试，覆盖前向传播、反向传播、损失计算、确定性验证

模型特性

• 加权有向消息传递图神经网络，用于聚合物性质预测
• 支持原子级和键级特征嵌入
• 基于PaddleMaterials标准API：forward(data, return_loss, return_prediction) → {"loss_dict", "pred_dict"}
• 已注册到 ppmat/models/__init__.py

测试

cd test && python -m pytest wd_mpnn/ -v
# 8 passed

参考链接

• 原始论文: https://doi.org/10.1021/acs.jcim.1c00975
• 原始仓库: https://github.com/coleygroup/polymer-chemprop (MIT License)
• Hackathon Issue: 【Hackathon 10th】开源贡献个人挑战赛 · 春节特别季 Paddle#77429

[paddle-bot]

Thanks for your contribution!

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

cloudforge1 seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}