InstaDeepAI/mace-organics

MACE

Reference

Ilyes Batatia, Dávid Péter Kovács, Gregor N. C. Simm, Christoph Ortner, and Gábor Csányi. Mace: Higher order equivariant message passing neural networks for fast and accurate force fields, 2023.

URL: https://arxiv.org/abs/2206.07697

How to Use

For complete usage instructions and more information, please refer to our documentation

Model architecture

Parameter	Value	Description
num_layers	2	Number of MACE layers.
num_channels	128	Number of channels.
l_max	3	Highest degree of spherical harmonics.
node_symmetry	3	Highest degree of node features kept after the node-wise power expansion of features.
correlation	2	Maximum correlation order.
readout_irreps	["16x0e","0e"]	Irreps for the readout block.
num_readout_heads	1	Number of readout heads.
include_pseudotensors	False	Whether to include pseudo-tensors.
num_bessel	8	Number of Bessel basis functions.
activation	silu	The activation function used in the non-linear readout block.
radial_envelope	polynomial_envelope	The radial envelope function.
symmetric_tensor_product_basis	False	Whether to use a symmetric tensor product basis.
atomic_energies	average	Treatment of the atomic energies.
avg_um_neighbors	None	Mean number of neighbors.

For more information about MACE hyperparameters, please refer to our documentation

Training

Training is performed over 220 epochs, with an exponential moving average (EMA) decay rate of 0.99. The model employs a MSE loss function with scheduled weights for the energy and force components. Initially, the energy term is weighted at 40 and the force term at 1000. At epoch 115, these weights are flipped. We use our default MLIP optimizer in v1.0.0 with the following settings:

Parameter	Value	Description
init_learning_rate	0.01	Initial learning rate.
peak_learning_rate	0.01	Peak learning rate.
final_learning_rate	0.01	Final learning rate.
weight_decay	0	Weight decay.
warmup_steps	4000	Number of optimizer warm-up steps.
transition_steps	360000	Number of optimizer transition steps.
grad_norm	500	Gradient norm used for gradient clipping.
num_gradient_accumulation_steps	1	Steps to accumulate before taking an optimizer step.

For more information about the optimizer, please refer to our documentation

Dataset

Parameter	Value	Description
graph_cutoff_angstrom	5	Graph cutoff distance (in Å).
max_n_node	32	Maximum number of nodes allowed in a batch.
max_n_edge	288	Maximum number of edges allowed in a batch.
batch_size	64	Number of graphs in a batch.

This model was trained on the SPICE2_curated dataset. For more information about dataset configuration please refer to our documentation

License summary

1. The Licensed Models are only available under this License for Non-Commercial Purposes.
2. You are permitted to reproduce, publish, share and adapt the Output generated by the Licensed Model only for Non-Commercial Purposes and in accordance with this License.
3. You may not use the Licensed Models or any of its Outputs in connection with:
   1. any Commercial Purposes, unless agreed by Us under a separate licence;
   2. to train, improve or otherwise influence the functionality or performance of any other third-party derivative model that is commercial or intended for a Commercial Purpose and is similar to the Licensed Models;
   3. to create models distilled or derived from the Outputs of the Licensed Models, unless such models are for Non-Commercial Purposes and open-sourced under the same license as the Licensed Models; or
   4. in violation of any applicable laws and regulations.