✨ Overview
In the current version, TACE implements models directly in Cartesian space using irreducible Cartesian tensor.
The framework supports prediction of a wide range of physical properties commonly required in atomistic simulations and materials modeling.
Supported conservative or direct properties include:
- ⚡ Energy
- 🧭 Forces
- 🧱 Stress
- 📦 Virials
- 🔋 Charges
- 🔭 Dipole Moment
- 🌐 Polarization
- 🧲 Polarizability
- 🧪 Born Effective Charges
- 🧭 Noncollinear Magnetic Forces
In addition, the architecture includes a series of interesting features such as partial Hessian training, spherical Bessel functions of arbitrary order ...
🧠 Universal Embedding
TACE introduces a universal embedding, allowing the model to incorporate additional invariants and equivariants.
Supported embeddings include:
- 🏷 Fidelity tags
- ⚡ Atomic charges
- 🌩 External electric fields
- 🧲 Initial noncollinear magnetic moments
This design enables richer physical conditioning and facilitates multi-fidelity workflows.
🏗 Foundation Models
This release introduces three foundation models trained in Cartesian space:
- TAVE-v1-OMat24-M
- TACE-v1-OAM-M
- TACE-v1-LES-REICO-5-PdAgCHO
🔧 Fine-Tuning Options
To support efficient adaptation of foundation models, TACE provides two fine-tuning strategies:
- LoRA (Low-Rank Adaptation)
- Freeze-based fine-tuning
These approaches enable flexible trade-offs between training efficiency, parameter updates, and computational cost.
🧩 Supported Calculators
TACE integrates with several widely used computational engines:
- ASE
- LAMMPS-ML-IAP
- TorchSim
🔌 Additional Plugins
TACE also provides optional extensions for enhanced physical modeling:
- Latent Ewald Summation (LES)
⚠️ Limitations of Cartesian Models
Although Cartesian TACE models can outperform spherical models at low truncation levels, these limitations constrain long-term scalability and efficiency.
- Lack of operator fusion libraries
- Exponential-level computational cost
- Limited optimization compared with spherical tensor frameworks
📌 Final Pure Cartesian Release
TACE v0.1.0 is the final version based purely on Cartesian tensor models.
Important notes:
⚠️ v0.1.0 will not remain compatible with future versions- Future development will transition to multi-backend frameworks
🚀 Roadmap for v0.2.0
Starting from v0.2.0, TACE will support multiple equivariant backends, including:
- cartnn
- e3nn
- equitorch
- openequivariance
- cuEquivariance
All model series will maintain consistent architecture and accuracy, but thanks to operator fusion libraries, spherical models will significantly outperform Cartesian implementations in both:
- ⚡ Speed
- 💾 Memory efficiency
Expected performance trend:
speed: otherTACE > cartTACE > e3nnTACE
memory: otherTACE > e3nnTACE > cartTACE
Additionally, new foundation model series will be released alongside the new backend system.
🔬 ACE Architecture Extensions
Beyond the standard frequency-domain coupling product basis and precomputed Cartesian product basis, TACE introduces spatial-domain product bases in both:
- Cartesian space
- Spherical space
This extension expands the flexibility of the Atomic Cluster Expansion architecture.
Current Limitation
Recent literature indicates that existing spatial ACE implementations lack antisymmetric interactions.
Support for antisymmetric interaction terms is currently under development and will be introduced in a future release.
🙏 Acknowledgements
We thank all contributors to this release: @xvzemin.
Full Changelog: v0.1.0...v0.1.0
