Roadmap

Current status (September 2024)

We set out to implement reference datasets and neural network potentials to provide a framework to assess the performance of the potentials within the same framework.

Reference implementations

• Datasets: SPICE1, SPICE2, ANI2x, ANI1x, QM9, PHALKETHOH
• Potentials: ANI2x, AimNet2, PhysNet, SchNet, PaiNN, SAKE, TensorNet
• Training routines: Training based on energies, forces, dipole moment, energy decomposition ($E_{\text{short}}$ + $E_{\text{elec}}$)

Trained models

• Base line models: Reference models trained on PHALKETHOH and SPICE2 dataset

To come

Q4 2024

• Add models based on alternative representations (Bessel functions, Spherical harmonics): DimNet++, So3krates
• develop API for openMM