MindScience 0.1.0

MindScience 0.1.0 Release Notes

Major Features and Improvements

MindSpore Elec

• Provide physics-driven and data-driven neural network for electromagnetic simulation
• Support CSG geometry model construction and CAD format data processing
• Include multiple scale filtering and dynamic adaptive weighted loss for improving performance
• Provide visualization tools for electromagnetic fields and scattering parameters

MindSPONGE

• Provide basic toolkits for molecular simulation, including MSA dataset, molecular pre-trained model(service on HUAWEI CLOUD), molecular dynamics。
• MSA dataset：Multiple Sequence Alignment Dataset for protein structure and function research
• Molecular Pre-trained Model：Trained with 1.7 billion compounds and its downstream tasks achieve SOTA
• Molecular Dynamics：Support basic MD functions，such as NPT, NVT, NVE and Minimization

MindChemistry

• Provide a high-entropy alloy composition design approach: Based on generation model and ranking model generating high-entropy alloy composition candidates and candidates' ranks, this approach constructs an active learning workflow for enabling chemistry experts to accelerate design of novel materials.
• Provide molecular energy prediction models: Based on equivariant computing library, the property prediction models NequIP and Allegro are trained effectively and infer molecular energy with high accuracy given atomic information.
• Provide an electronic Structure Prediction model: We integrate the DeephE3nnn model, an equivariant neural network based on E3, to predict a Hamiltonian by using the structure of atoms.
• Provide a crystalline material properties prediction model: We integrate the Matformer model, based on graph neural networks and Transformer architectures, for predicting various properties of crystalline materials.
• Provide an equivariant computing library: We provide basic modules such as Irreps, Spherical Harmonics as well as user-friendly equivariant layers such as equivarant Activation and Linear layers for easy construction of equivariant neural networks.

Contributors

Thanks goes to these wonderful people:

yufan, gaoyiqin, wangzidong, yangkang, lujiale, shibeiji, liuhongsheng, liyang, wengbingya, chuhaotian, huangxiang, wangmin, niningxi, zhangxinfeng, yujialiang, qianjiahong, chenmengyun, yanglijiang, yangyi, huangyupeng, xiayijie, zhangjun, linxiaohan, chendiqing, gongyue, gengchenhua, linghejing, yanchaojie, suyun, wujian, caowenbin

Contributions of any kind are welcome!