Mind Blowing New Foundational Model in Materials Chemistry! 🧪🤖💥
So essentially,
"MACE Foundational Model is evolving our understanding of developing chemicals at an atomic level!"
Paper: A foundation model for atomistic materials chemistry (105 pages)
Researchers from multiple institutes in the UK, Germany, and the USA are interested in more accurate simulations for different material interactions. Specifically, they are interested in enabling efficient simulations of the ab initio quality.
Here’s the background:
Atomic-scale simulations are currently based on density functional theory which is hard to computationally model and scale
Existing analytical models based on inter-atomic force fields are relatively limited in accuracy and relevance
The applications of this technology involve water system predictions, metal-organic frameworks as well as inorganic crystals and elemental lattices
Ok, so what was the research?
The researchers use MACE architecture, which is a message-passing graph tensor network based on the MPtrj dataset. Here are some details of the training:
The dataset included 150k inorganic crystals from the Materials Project database
The data was processed into a format that was compatible with the MACE-MP-0
The loss function for the model is the mean absolute error (MAE) between the predicted and target values of the energy, forces, and stress
In each batch updating step, the weighted sum of Huber losses (251) of energy, forces, and stress incurred by all structures in a batch are averaged and back-propagated into the neural networks
MACE-MP-0 was shown to stably run molecular dynamics simulations across diverse chemical systems, including solids, liquids, gases, and chemical reactions. It could predict phonon spectra, calculate activation energies for point defect and dislocation motion, simulate solvent mixtures, combust hydrogen gas, model a complete rechargeable battery cell, and more.
And what’s next?
Currently, the model struggles to describe intermolecular interactions and high-pressure simulations which can lead to qualitative and quantitative errors. As the model is trained on data generated using DFT calculations, it has limitations in accuracy. Refining the model using higher-level electronic structure theory data can further improve its accuracy.
Despite its limitations, there is incredible promise in this foundational model!
The MACE model can be used to predict the properties of materials, such as their stability, structure, and reactivity. This information can be used to identify new materials with desirable properties such as to identify new drugs or drug targets, design more efficient and environmentally friendly chemical processes, storing and generating renewable energy, and a wide variety of other applications, such as manufacturing, aerospace, and defense.
So essentially,
"MACE Foundational Model is evolving our understanding of developing chemicals at an atomic level!"