Release notes

Release notes#

[v2.8.2] - 2025-12-01#

This release focuses on integrating uncertainty quantification into the Metatomic potential interface and modernizing the build and dependency infrastructure.

✨ New Features#

  • Metatomic Uncertainty: Added support for checking per-atom energy uncertainty in Metatomic potentials.

    • Introduced the uncertainty_threshold parameter (default: 0.1, corresponding to ~100 meV/atom).

    • If the model supports the energy_uncertainty output (e.g., PET-MAD), EON will now warn if atoms exceed this threshold.

    • The variance is now automatically populated using the mean of these per-atom uncertainties if available.

  • Pixi Build System: Added new environments (rel, eon) and tasks to pixi.toml to streamline the release process and simplify local builds.

🚀 Enhancements#

  • Device Selection: The Metatomic potential now utilizes the metatomic_torch::pick_device API for more robust device selection (CPU/GPU) based on model capabilities.

  • Model Loading: Updated the internal loading mechanism to use metatensor_torch::Module for better compatibility with recent Metatensor versions.

📦 Dependencies#

  • PyTorch: Updated to v2.8.

  • Metatomic/Metatensor: Updated dependencies to metatomic-torch >= 0.1.7 and metatensor-torch >= 0.8.2 to support the new uncertainty APIs.

  • Pydantic: Pinned specific versions to resolve linting and compatibility issues in the build environment.

Contents