Machine Learning-enhanced Quantum Chemistry

Figure: A Model Structure-Neural Network Process

Researchers from Los Alamos National Laboratory proposed incorporating more quantum mechanics mathematics into the structure of machine learning predictions in a new study published in Proceedings of the National Academy of Sciences. The machine learning model predicts an effective Hamiltonian matrix based on the specific positions of atoms within a molecule, which describes the various possible electronic states and their associated energies.

The machine learning-based approach makes predictions at a much lower computational cost than traditional quantum chemistry simulations. It allows for quantitatively precise predictions of material properties, interpretable insight into the nature of chemical bonding between atoms, and can be used to predict other complex phenomena, such as how the system will respond to perturbations like light-matter interactions. The method also outperforms traditional machine learning models in terms of accuracy and transferability, i.e., the model's ability to make predictions that extend far beyond the data that served as the basis for its training.

Starting with basic scientific theories, quantum mechanics equations provide a roadmap for predicting chemical properties. When used to predict behavior in large systems, these equations can quickly become too expensive in terms of computer time and power. Machine learning appears to be a promising method for speeding up such large-scale simulations. The use of machine learning to predict chemical properties has the potential to lead to significant technological advancements, with applications ranging from cleaner energy to faster pharmaceutical drug design. This is a very active area of research, but most existing approaches to machine learning model design are simple and heuristic.

The researchers demonstrated in their study that machine learning models can mimic the basic structure of nature's fundamental laws. These laws can be extremely difficult to directly simulate. In a wide range of chemical systems, the machine learning approach enables predictions that are simple to compute and accurate.

The improved machine learning model can predict a wide range of molecule properties quickly and accurately. These approaches perform very well on important benchmarks in computational chemistry and demonstrate how deep learning methods can improve further by incorporating more experimental data. The model is also capable of predicting excited state dynamics, or how systems behave when their energy levels are elevated. This tool represents a quantum chemistry breakthrough. It will help scientists understand the reactivity and excited states of new molecules.