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 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 on which it is trained and allows for dramatic computational savings by indirectly targeting the outputs of well-converged calculations while using a parametrization 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 primarily addresses the problem of efficiently predicting molecular electronic excited states by combining machine learning (ML) methods with quantum mechanics (QM) calculations. Specifically, the research team developed an integrated modeling approach that uses a symmetry-adapted machine learning model to predict an effective Hamiltonian, which is trained to reproduce electronic excitations obtained from quantum mechanical calculations. This approach allows for predictions on larger and more complex molecules than those in the training data, while significantly reducing computational costs. Key points of the paper include: - Proposing an indirect learning framework where the machine learning model is used to predict elements of a single-particle effective Hamiltonian, which are then used to compute molecular orbital energy levels and other electronic structure properties. - Discussing several different training strategies, including learning directly for Hamiltonian matrix elements, learning only for molecular orbital energy levels under the minimal basis set, and learning considering both energy levels and Löwdin charges. - Demonstrating that training with larger basis sets as targets, even when the model architecture is still based on the minimal basis set, can achieve high accuracy. - Showing that the effective Hamiltonian obtained from machine learning can be further used to compute electronic excited state energies with good accuracy. - Indicating that the method is not only applicable to small molecules in the training set but also generalizes well to larger molecular systems, including long-chain polyenes and aromatic compounds. - Finally, illustrating how the hybrid model can be combined with advanced simulation techniques to predict fine quantum mechanical effects, such as the vibrational spectrum of an anthracene molecule, at low cost. In summary, this work presents a novel approach that enhances the transferability and interpretability of machine learning models by combining data-driven techniques with physics-based approximations, while maintaining high accuracy and computational efficiency.

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

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

Machine learning the electronic structure of matter across temperatures

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

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

Synergy of semiempirical models and machine learning in computational chemistry

Machine learning for molecular simulation

Machine Learning Frenkel Hamiltonian Parameters to Accelerate Simulations of Exciton Dynamics

Resolution-vs.-Accuracy Dilemma in Machine Learning Modeling of Electronic Excitation Spectra

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

Machine Learning Exciton Hamiltonians in Light-Harvesting Complexes

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

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

Modeling electronic response properties with an explicit-electron machine learning potential

Machine learning and excited-state molecular dynamics