Machine Learning Potentials (MLPs) for Accelerated Battery Electrolyte Molecular Dynamics
Accelerate Battery Discovery: From Quantum Accuracy to Million-Atom Simulations with AI
Skills you will gain:
About Program:
Density Functional Theory (DFT) offers quantum-level accuracy but is fundamentally bottlenecked by its high computational cost—limiting simulations to a few hundred atoms. To discover the next generation of solid-state batteries, researchers must bridge the gap between quantum accuracy and macro-scale realities.
This workshop provides a comprehensive blueprint for training and deploying Machine Learning Potentials (MLPs) like SNAP, MACE, and CHGNet. By training AI models on DFT data, you will learn how to scale your simulations to tens of thousands of atoms, accelerating solid-state electrolyte discovery and generating publication-ready insights in a fraction of the time.
Aim: Learn to leverage Machine Learning Potentials (MLPs) to scale molecular dynamics simulations from hundreds to hundreds of thousands of atoms.
Gain hands-on experience in training, validating, and deploying AI-driven models for battery electrolyte discovery.
Program Objectives:
- Understand the limits of DFT and AIMD and how AI bridges the scale gap.
- Train and validate Machine Learning Potentials using MLP frameworks.
- Deploy MLPs for large-scale molecular dynamics simulations.
- Analyze battery performance metrics: ionic conductivity, MSD, and diffusion coefficients.
- Integrate AI workflows into publication-ready and reproducible pipelines.
What you will learn?
Day 1: Quantum to AI – Data Curation & Interatomic Potential Fitting
- Focus: Data engineering for materials and bridging DFT data with Machine Learning frameworks.
- The computational limits of ab initio Molecular Dynamics (AIMD) in battery interface modeling.
- Architectural overview of modern MLPs: Descriptor-based (SNAP), Graph Neural Networks (CHGNet), and Equivariant GNNs (MACE).
- Data preprocessing workflows: Extracting energies, forces, and stresses using the Atomic Simulation Environment (ASE).
- 🚀 Hands-on Lab (Google Colab): Dataset to Potential
Import a raw DFT dataset for a lithium-ion solid electrolyte, preprocess the structural configurations using ASE, and initialize an equivariant machine learning model for training.
Day 2: High-Performance Training with DeepMD-kit
- Focus: Training, optimizing, and validating neural network potentials for production-grade accuracy.
- DeepMD-kit Architecture: Multi-task learning routines for mapping atomic coordinates to total energy and forces.
- Hyperparameter optimization: Balancing loss function weights ($E$ vs. $F$ vs. $V$) and managing learning rate decay.
- Error analysis for high-impact publishing: Evaluating Root Mean Square Error (RMSE) against validation benchmarks.
- 🚀 Hands-on Lab (Google Colab): Training the Neural Network
Configure a DeepMD-kit training pipeline using a JSON framework, execute a multi-epoch training run on a cloud GPU, and analyze the loss convergence behavior.
Day 3: Large-Scale Production MD Simulation via LAMMPS
- Focus: Deploying trained MLPs into scalable, real-world molecular dynamics production runs.
- Interfacing trained ML models into LAMMPS for high-performance computing.
- Configuring large-scale ensembles (NVT/NPT) for complex solid-state electrolyte interfaces.
- Post-processing trajectories: Extracting critical battery metrics including Lithium-ion Conductivity and Mean Squared Displacement (MSD).
- 🚀 Hands-on Lab (Google Colab): The Million-Atom Simulator
Deploy your trained ML potential directly into a LAMMPS execution block, simulate a large-scale solid electrolyte system, and calculate the target ionic diffusion coefficients.
Mentor Profile
Fee Plan
Get an e-Certificate of Participation!
Intended For :
- Researchers, PhD scholars, and postdocs in Materials Science, Computational Chemistry, Physics, or Chemical Engineering.
- Industry R&D professionals working on battery materials, electrochemistry, or computational modeling.
- Individuals with basic Python and molecular modeling knowledge seeking hands-on AI-driven simulation skills.
Career Supporting Skills
Program Outcomes
- Gain hands-on experience in preprocessing DFT datasets for machine learning potential training.
- Train and validate MLPs (SNAP, MACE, CHGNet, DeepMD-kit) for atomic-scale simulations.
- Scale molecular dynamics simulations from hundreds to hundreds of thousands of atoms.
- Analyze battery-relevant properties such as ionic conductivity, mean squared displacement (MSD), and diffusion coefficients.
- Apply AI workflows to accelerate research pipelines and generate publication-ready results.
- Deploy trained potentials in LAMMPS for high-performance, large-scale production runs.
- Integrate cloud-based GPU computing via Google Colab to run intensive simulations without local HPC infrastructure.