How to train your first ML Interatomic Potential using NequIP? [TUTORIAL 1]

This YouTube tutorial by me (Manas Sharma) demonstrates training a NequIP machine learning model. The tutorial uses a silicon crystal's atomic molecular dynamics trajectory data, available for download, to train the model on Kaggle. I explain NequIP's functionality, focusing on its ability to accurately predict energies and forces in chemical systems by exploiting 3D symmetries. The tutorial provides step-by-step instructions and code, ultimately resulting in a deployable model with highly accurate predictions. Here are some key points covered in the video: * *Training a Model from Scratch* The video demonstrates how to train a NequIP machine learning model from the very beginning. * *No powerful local computer needed* The tutorial uses Kaggle to train the model online, meaning you don't need a GPU on your local machine. * *Overview of the NequIP model* The video explains that NEQUIP stands for Neural Equivariant Interacting Potentials. It discusses the neural network, equivariant, and interacting potential aspects of the model. The model is designed for chemical systems, predicting energies and forces, and can be used for molecular dynamics simulations. The model is a message passing graph neural network, using atoms as nodes and pairs of atoms as edges . * *Input file* The video uses an extended XYZ file format as the input file, which includes the number of atoms, atomic positions, and atomic forces. The data is from an AIMD (Ab Initio Molecular Dynamics) trajectory. * *Training and Analyzing the Model* The video explains how to use a config YAML file to set the parameters for the graph neural network, such as the radial cutoff, number of layers, and the order of the features of the network . The video also shows how to analyze metrics during the training . * *Testing and Deploying the Model*: The video demonstrates how to evaluate the model's accuracy using test data and then deploy the trained model to be used in other simulation programs such as LAMMPS. *Timestamps:* * 0:00:00 - Introduction * 0:00:13 - Channel subscription request * 0:00:43 - Introduction to NequIP * 0:01:47 - Desirable qualities of a machine learning model * 0:03:59 - NequIP architecture * 0:05:57 - Training the model using a silicon crystal trajectory * 0:07:29 - Using Kaggle to train the model * 0:09:54 - Downloading trajectory file * 0:10:28 - Creating the config YAML file * 0:17:11 - Training the model * 0:19:04 - Evaluating test error * 0:21:26 - Deploying the model * 0:22:18 - Downloading the files * 0:23:13 - Analyzing the results * 0:24:25 - Conclusion *IMPORTANT LINKS:* NequIP Paper: https://www.nature.com/articles/s41467-022-29939-5 NequIP GitHub Repo: https://github.com/mir-group/nequip Jupyter Notebook for Kaggle: https://www.bragitoff.com/wp-content/uploads/2025/01/NequIP-Tutorial-YouTube-Silicon.ipynb Kaggle notebook link: https://www.kaggle.com/code/ducktape07/nequip-tutorial-youtube-silicon Jupyter Notebook as PDF: https://www.bragitoff.com/wp-content/uploads/2025/01/NequIP-Tutorial-YouTube-Silicon.pdf Trajectory file for training: https://www.bragitoff.com/wp-content/uploads/2025/01/sitraj.xyz Final results directory: https://www.bragitoff.com/wp-content/uploads/2025/01/training_results_kaggle_si.zip Other useful resources from authors of nequip: https://deepnote.com/app/shuai-jiang-c648/NequIP-Tutorial-Duplicate-b4c9a903-0586-4587-a353-85181f017467 https://www.youtube.com/watch?v=oePOO8bN7Co https://www.youtube.com/watch?v=ZR1NTBPBDOo https://www.youtube.com/watch?v=-mRl5Uk8IWk ---------------------------------------------------------------------- Connect with Me: 💻Website: https://manas.bragitoff.com LinkedIn: https://www.linkedin.com/in/manassharma07/ Twitter: https://x.com/ManasSharma07