Compressing physics with an autoencoder: Creating an atomic species representation to improve machine learning models...

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Accurate prediction of material properties is crucial for discovering new materials.
• Current machine learning models for materials often face limitations in scalability and the number of elements they can handle.
• Developing efficient feature representations for atomic species is key to advancing materials modeling.

Purpose of the Study:

• To create a novel, compressed vector representation of atomic species identity.
• To improve the accuracy and scalability of machine learning models for predicting material properties, specifically formation energies.
• To enable the application of high-dimensional neural network potentials (HD-NNPs) to a wider range of elements and complex chemical processes.

Main Methods:

• Utilized an autoencoder to compress physical properties into a vector quantity representing atomic identity, termed elemental modes.
• Trained neural networks using elemental modes as feature vectors to predict formation energies of elpasolite compounds.
• Integrated elemental modes with geometric features for HD-NNPs, extending their capability to numerous atomic species and alchemical intermediate states.

Main Results:

• Elemental modes significantly improved the accuracy of predicting elpasolite formation energies compared to previous methods.
• The new approach overcomes the scaling limitations of traditional HD-NNPs, allowing for models with up to 11 atomic species (H, C, N, O, F, P, S, Cl, Se, Br, I).
• Demonstrated the ability to define feature vectors for alchemical intermediate states, facilitating free energy calculations in systems with bond breaking/forming.

Conclusions:

• Elemental modes offer a powerful and efficient feature representation for atomic species in machine learning.
• This advancement expands the applicability of HD-NNPs to a larger chemical space and complex reaction pathways.
• The method opens new avenues for alchemical free energy calculations, crucial for understanding chemical transformations.