MatPES.ai: A Foundational Dataset for Materials Science AI

MatPES.ai is a cutting-edge, open-access dataset designed to advance machine learning interatomic potentials (MLIPs) in materials science by providing high-accuracy, comprehensive, and rigorously validated potential energy surface (PES) data. Developed by the Materials Virtual Lab and the Materials Project, it leverages static DFT calculations with stringent convergence criteria and includes data from both the PBE and r2SCAN functionals, ensuring broad chemical coverage and improved transferability across diverse materials. The platform supports researchers and developers through interactive data exploration, command-line tools, Python packages, and Jupyter notebook tutorials, enabling efficient data retrieval, model training, and fine-tuning. It is seamlessly integrated into the MatML ecosystem, powering advanced architectures like M3GNet, CHGNet, and TensorNet, and is optimized for use with MatCalc for rapid property prediction. With over 400,000 structures from 300K MD simulations and detailed performance benchmarks, MatPES.ai sets a new standard for foundational materials data in AI-driven discovery.

Key Features:

• Comprehensive PES Dataset: Near-complete periodic table coverage with ~400,000 structures from 300K MD simulations.
• High-Accuracy DFT Calculations: Computed using stringent convergence criteria with both PBE and r2SCAN functionals for enhanced reliability.
• Dual Functional Support: Includes data for PBE and r2SCAN meta-GGA functionals, with r2SCAN offering improved accuracy across diverse bonding types.
• 2-Stage DIRECT Sampling: Ensures diverse and representative atomic configurations for robust model training.
• Interactive Data Explorer: Visualize material properties via elemental heatmaps and property distributions using Plotly; filter by chemical system with an interactive periodic table.
• Seamless Integration with MatML Ecosystem: Works with MatGL, maml, and MatCalc for model training, inference, and property computation.
• Python Package & CLI Tools: Install via pip install matpes and use command-line tools for downloading and filtering data (e.g., matpes download pbe, matpes data -i ... --chemsys Fe-O -o ...).
• MongoDB Support: The matpes.db module enables efficient database creation and querying for large-scale data management.
• Pre-Trained Foundation Potentials: Models available in M3GNet, CHGNet, and TensorNet architectures with standardized naming (e.g., TensorNet-MatPES-PBE-v2025.1-PES).
• Jupyter Notebook Tutorials: Step-by-step guides for data loading, model training, and fine-tuning.
• Performance Benchmarks: Rigorous evaluation of MLIPs using paired t-tests and ASE-compatible calculators; includes equilibrium, near-equilibrium, and molecular dynamics testing.
• VASP Input Generation: Tool to generate VASP inputs compatible with MatPES workflows using pymatgen.

Pricing: MatPES.ai is offered as a free and open resource for the scientific community, with no cost for access or usage.

Conclusion: MatPES.ai is an essential, high-quality, and accessible foundation for AI-driven materials discovery, empowering researchers with reliable, scalable, and well-documented data and tools to train universal machine learning potentials and accelerate innovation in computational materials science.