In this release, we have continued to develop our ML force field framework using Moment Tensor Potentials (MTPs). 1000-10,000x shorter computational time vs DFT enables ab initio accuracy for large system sizes and time scales greatly exceeding those accessible to DFT.
- Available MTP library with pre-trained MTPs for a range of bulk materials and interfaces.
Updates to the Core MTP Framework
- Significantly improved accuracy for multi-element systems using more a flexible design of the potential coefficients
- New extrapolation grade algorithm for active learning: Query-by-Committee
- The most important MTP training workflows are now accessible in the new Workflow Builder in the Nanolab GUI for fast and easy setup.
- Built-in MTP analyzer to plot correlation between MTP and reference DFT energy, forces, and stress from Data Tool in Nanolab
- Training protocol updates (generating training data for alloy materials and improvements for interface training protocol)
- Improvements to MTP training functionality
- CrystalPropertyValidation analysis object
Pretrained MTP library
- New pre-trained MTP models
- Re-trained MTP models for HKMG stack and interfaces, with added gate metals Ru, Sc, and their interface with HfO2
- Re-trained TiNAlO potential with improved accuracy
Multilayer Stack Generation using the Multilayer Builder GUI
- New MTJ builder to build stack configurations for magnetic RAM (MRAM) applications
- Improvements in features and quality of the generated structures from the HKMG builder
- General multilayer builder improvements, specifying displacement vectors for better aligning of crystal interfaces
- Tutorials to explain how to build HKMG and MRAM stacks