Abstract:Data-driven techniques are increasingly used to replace electronic-structure calculations of matter. In this context, a relevant question is whether machine learning (ML) should be applied directly to predict the desired properties or be combined explicitly with physically-grounded operations. We present an example of an integrated modeling approach, in which a symmetry-adapted ML model of an effective Hamiltonian is trained to reproduce electronic excitations from a quantum-mechanical calculation. The resulting model can make predictions for molecules that are much larger and more complex than those that it is trained on, and allows for dramatic computational savings by indirectly targeting the outputs of well-converged calculations while using a parameterization corresponding to a minimal atom-centered basis. These results emphasize the merits of intertwining data-driven techniques with physical approximations, improving the transferability and interpretability of ML models without affecting their accuracy and computational efficiency, and providing a blueprint for developing ML-augmented electronic-structure methods.

What problem does this paper attempt to address?

The paper aims to address the problem of efficiently and accurately predicting the electronic excited states of molecules by combining machine learning (ML) techniques with quantum mechanics (QM) calculations. Specifically, the researchers propose an integrated modeling approach that trains a symmetry-adapted machine learning model to simulate an effective Hamiltonian matrix, reproducing the electronic excited states obtained from quantum mechanical calculations. This approach not only allows for predictions on molecules larger and more complex than those in the training data but also significantly reduces computational costs. The main contributions of the paper include: 1. **Proposing a new hybrid machine learning architecture**: This architecture combines physical constraints with machine learning by constructing an effective Hamiltonian matrix with the correct symmetry properties to predict physical quantities such as molecular orbital energies and atomic charges. This method enhances the model's transferability and interpretability while maintaining accuracy. 2. **Achieving significant computational efficiency improvements**: Compared to traditional quantum mechanical calculation methods, this approach offers significant computational efficiency advantages in predicting molecular orbital energies and electronic excitation energies, especially when dealing with large molecules. 3. **Demonstrating good generalization ability of the model**: The researchers show that the model not only accurately predicts the electronic structure properties of molecules in the training set but also accurately predicts the electronic excitation energies of unseen large molecules, indicating good generalization ability. 4. **Exploring the impact of different training strategies**: The study compares the effects of various training strategies, including directly predicting Hamiltonian matrix elements, indirectly predicting through molecular orbital energies, and combining molecular orbital energies and atomic charges. It finds that combining multiple targets can achieve more balanced and accurate prediction results. In summary, the method proposed in this paper provides a powerful tool for developing efficient and accurate electronic structure calculation methods and demonstrates the value of combining data-driven techniques with physical principles.

Electronic excited states from physically-constrained machine learning

Electronic Excited States from Physically Constrained Machine Learning

Machine Learning of Molecular Electronic Properties in Chemical Compound Space

Machine Learning for Electronically Excited States of Molecules

Molecular excited states through a machine learning lens

Machine Learning in Electronic Quantum Matter Imaging Experiments

Perspective on integrating machine learning into computational chemistry and materials science.

Using Machine Learning for Scientific Discovery in Electronic Quantum Matter Visualization Experiments.

Electrostatic Embedding Machine Learning for Ground and Excited State Molecular Dynamics of Solvated Molecules

Synergy of semiempirical models and machine learning in computational chemistry

Machine learning the electronic structure of matter across temperatures

Machine learning for molecular simulation

Machine Learning Frenkel Hamiltonian Parameters to Accelerate Simulations of Exciton Dynamics

Machine Learning Exciton Hamiltonians in Light-Harvesting Complexes

Machine learning and excited-state molecular dynamics

Machine learning model for non-equilibrium structures and energies of simple molecules

Incorporating electronic information into Machine Learning potential energy surfaces via approaching the ground-state electronic energy as a function of atom-based electronic populations

Multi-task learning for molecular electronic structure approaching coupled-cluster accuracy

Machine Learning in QM/MM Molecular Dynamics Simulations of Condensed-Phase Systems

Machine learning for accuracy in density functional approximations

Quantum Machine Learning for Chemistry and Physics