Data-driven quantum embedding enables faster simulations and reduces computational bottlenecks in DFT

Materials simulations often struggle with strongly correlated systems in which electrons interact in complex ways, but new approaches to quantum embedding promise to overcome these limitations. Samuele Juri of the Flatiron Institute, Hasanat Hasan of the Rochester Institute of Technology, and Benedict Cross, also of the Flatiron Institute, along with Marius S. Frank, Tsunhan Li, Olivier Gingras, and others demonstrated a method to compress the computations required to describe the interactions of a piece of material with its surroundings. Their work introduces a linear model that uses principal component analysis to learn the most important quantum states, significantly reducing the amount of computation required to solve complex problems. The researchers validated the method both on a simplified model system and on the notoriously difficult element plutonium, achieving significant speedups while maintaining accuracy, paving the way for routine simulations of highly correlated materials at comparable computational costs to standard methods.

Ghost Guttwiller approximation and correlated electrons

This is a comprehensive overview of research papers and preprints on strongly correlated electron systems and quantum many-body physics, including applications in machine learning and quantum computing. Key themes include Ghost-Gutwiller approximation (GGA), a technique for approximating solutions to complex many-body problems. It is used both as a standalone technique and as a component of more advanced techniques. Researchers have rigorously evaluated the accuracy of GGA and explored improvements such as bus size optimization and combination with density functional theory (DFT), as well as demonstrating its use for efficient screening of correlated materials. Dynamic mean field theory (DMFT) serves as the underlying method, often utilizing GGA as an impurity solver.

We also consider the LDA+U method, a common approach for handling strong correlations in DFT. Several studies have addressed the sign problem in quantum Monte Carlo (QMC) simulations, a major obstacle in accurately modeling many-body systems, and revisited the Mott transition of excitons, a key phenomenon in strongly correlated materials. A rapidly growing field focuses on neural network quantum states (NNQS) representing many-body wave functions. The fundamental model is studied for interatomic potentials and quantum states, aiming for generalizability across different Hamiltonians, and in particular NNQS is used as an impurity solver within a quantum embedding scheme.

Machine learning also improves the accuracy of variational quantum eigensolver (VQE) calculations and reduces sign problems for QMC with basis rotations. Several papers focus on developing machine learning interatomic potentials for materials simulation, particularly battery electrolytes and catalysts, utilizing meta-learning and multi-fidelity training to increase efficiency and accuracy. Researchers combine GGA and machine learning to screen materials for strong correlation effects. Quantum computing and hybrid quantum-classical approaches are also prominent, where variational quantum imaginary time evolution (VQITE) is used to prepare the ground state.

By combining GGA and quantum computing, the strengths of both approaches are leveraged and a quantum-assisted GGA analysis is proposed. The fermion-Gaussian circuit represents a quantum impurity model. Studies of crossmagnetism, new magnetic states, and spin-promoting effects in catalytic ammonia synthesis are ongoing in parallel with consideration of the role of Hund coupling in the strong correlation. The most exciting research is being conducted at the intersection of classical methods, machine learning, and quantum computing, with a focus on materials discovery, reducing errors, and developing fundamental models for materials science.

Principal component analysis accelerates quantum embedding

Scientists have developed a new computational method to address the limitations of traditional density functional theory in simulating strongly correlated materials. The research team addressed computational bottlenecks associated with quantum embedding (QE) theory, which maps complex systems into fragments that interact with their environment. Instead of directly solving a large-scale embedded Hamiltonian, they pioneered a linear model that utilizes principal component analysis (PCA) to compress the state space required for its solution, reducing the problem to finding ground states in a smaller data-driven variational subspace. The method employs an active learning scheme to learn an optimal variational subspace from the ground states of an embedded Hamiltonian, transforming the iterative solution process into a deterministic eigenvalue problem.

Experiments using the Ghost-Gutwiller approximation and the three-orbit Hubbard model demonstrate the effectiveness of the method and reveal that variational spaces trained on Bethe lattices can accurately predict the behavior of both square and cubic lattices without further training. This transferability represents a significant advance and eliminates the need to iteratively train between different lattice structures. Applying this method to plutonium, they succeeded in reproducing the energies of all six crystalline phases in a single variational space. The results show that the computational cost is significantly reduced and the time required to solve the embedded Hamiltonian is reduced by orders of magnitude. This work provides a practical path towards high-throughput first-principles simulations of strongly correlated materials with computational costs comparable to standard density functional theory, opening new avenues in materials discovery and design. The team plans to extend this data-driven framework to even larger embedded Hamiltonian problems by leveraging data generated from advanced solvers such as matrix product states and neural quantum states.

PCA efficiently compresses correlated material simulations

Scientists have achieved a breakthrough in computational materials science by developing a new method to simulate strongly correlated materials, overcoming a major bottleneck in traditional density functional theory (DFT) calculations. This work focuses on embedding theory to address the limitations of DFT for complex systems, which has traditionally suffered from high computational costs due to iterative solution of large-scale embedding Hamiltonians. In this study, we introduce a linear model using principal component analysis (PCA) to compress the information needed to solve this Hamiltonian and significantly reduce the amount of calculation. The research team demonstrated that the variational space learned from the ground state of an embedded Hamiltonian on a Bethe lattice is transferable to both square and cubic lattices without the need for further training.

This transferability is an important achievement and simplifies the process of applying this method to various material structures. Experiments using a three-orbit Hubbard model reveal that the PCA-based approach significantly reduces the costs associated with solving embedded Hamiltonians, paving the way for more efficient simulations. Measurements confirm that the method accurately captures the energies of all six crystalline phases of plutonium using a single variational space. This breakthrough reduces the computational cost of embedded Hamiltonian solutions by orders of magnitude.

The method includes a data-driven active learning scheme that learns a reduced variational subspace from ground states, transforming a complex problem into a deterministic eigenvalue problem within this small space. Tests demonstrate the robustness of our approach, which dynamically determines the dimensionality of the variational space based on an accuracy threshold, ensuring a highly accurate compressed representation of the data. This research establishes a practical path to high throughput. Abu Initio Strongly correlated materials simulation has the potential to enable material discovery and design at a cost comparable to standard DFT calculations.

Embedding theory simplifies correlated material simulations

Scientists have developed a new method to improve the accuracy and efficiency of simulations involving strongly correlated materials, a type of material in which electrons interact in complex ways. Current computational techniques based on Kohn-Sham density functional theory often struggle to handle these materials and yield inaccurate results. In this study, we introduce a linear model within an embedded theory framework that compresses the computational space required to describe the interactions between different parts of the material, significantly reducing the computational load. This method utilizes principal component analysis to identify and retain only the most important information needed to solve complex equations.

The team demonstrated the effectiveness of this approach using both model systems and real materials, particularly plutonium. In model calculations, we successfully applied the variational space learned for one crystal structure to other crystal structures, demonstrating its wide applicability. Importantly, the researchers achieved a significant reduction in computational costs when simulating all six crystalline phases of plutonium, demonstrating a practical path toward high-throughput simulations of strongly correlated materials at a cost comparable to standard density functional theory. This progress is expected to accelerate materials discovery and improve our understanding of complex electronic phenomena. The authors also acknowledge this.