Publications by authors named "Jakub Martinka"
J Chem Phys
November 2024
Article Synopsis
- Machine learning (ML) for predicting vector or tensor properties requires maintaining proper invariance with molecular rotation, unlike energy prediction which is simpler due to its scalar nature.
- The proposed "rotate-predict-rotate" (RPR) technique involves three steps: rotating the molecule to align with its principal axes, predicting the vector property in that orientation, and then transforming the prediction back to the original orientation.
- This RPR approach ensures covariance for vector properties and can extend to tensors, allowing for rapid training of accurate ML models across numerous molecular configurations.
