Abstract:Variational ab-initio methods in quantum chemistry stand out among other methods in providing direct access to the wave function. This allows in principle straightforward extraction of any other observable of interest, besides the energy, but in practice this extraction is often technically difficult and computationally impractical. Here, we consider the electron density as a central observable in quantum chemistry and introduce a novel method to obtain accurate densities from real-space many-electron wave functions by representing the density with a neural network that captures known asymptotic properties and is trained from the wave function by score matching and noise-contrastive estimation. We use variational quantum Monte Carlo with deep-learning ansätze (deep QMC) to obtain highly accurate wave functions free of basis set errors, and from them, using our novel method, correspondingly accurate electron densities, which we demonstrate by calculating dipole moments, nuclear forces, contact densities, and other density-based properties.

What problem does this paper attempt to address?

### Problems the Paper Attempts to Solve This paper aims to address the technical challenge of extracting accurate electron density from multi-electron wave functions in quantum chemistry. Specifically: 1. **Advantages of Direct Access to Wave Functions**: Variational ab initio methods (such as Variational Quantum Monte Carlo, VMC) can directly access wave functions, theoretically allowing the extraction of any observable of interest directly from the wave function, not just energy. However, in practice, these extraction processes are often technically difficult and computationally expensive. 2. **Limitations of Existing Methods**: - **Real-Space Wave Functions**: When using real-space wave functions (such as those produced by VMC), it is necessary to carefully design low-variance estimators and sample each point where density needs to be determined individually, which limits resolution and further applications. - **Second Quantized Wave Functions**: When using second quantized wave functions (such as those produced by Configuration Interaction, CI), the accuracy of the basis set is severely limited, and basis set extrapolation techniques are not applicable to electron density. - **Non-Variational Methods**: Methods like Coupled Cluster and Møller–Plesset can improve the accuracy of total energy, but the methods for extracting observables from them are neither unique nor straightforward. 3. **Importance of Electron Density**: The one-electron density \( \rho(r) \) succinctly describes many electronic properties of matter, forms the foundation of Density Functional Theory, and is crucial for bond classification and the calculation of other derived observables (such as electrostatic multipole moments, nuclear forces, contact density, etc.). ### Solution To address the above issues, the paper introduces a new method—Neural Network Electron Real-space Density (NERD), which achieves high-precision electron density extraction through the following steps: 1. **Using Deep Learning Wave Functions**: Employing Variational Quantum Monte Carlo (VMC) combined with deep learning wave functions (Deep QMC) to obtain high-precision wave functions without basis set errors. 2. **Neural Network Representation of Density**: Representing electron density as a neural network model that incorporates known asymptotic properties and is trained from the wave function using score matching and noise contrastive estimation. 3. **Asymptotic Correctness**: The NERD model is designed to be asymptotically correct both near and far from the atomic nuclei. 4. **Efficient Computation**: Once trained, the NERD model can provide density values at any position in space with very low computational cost. ### Experimental Validation The paper validates the effectiveness of the NERD method through the following experiments: 1. **Comparison on Small Molecules**: Comparing the density obtained by the NERD method with other methods such as Full Configuration Interaction (FCI) on small molecules, showing that the NERD method captures the same overall features but performs better in asymptotic behavior. 2. **Calculation of Derivative-Related Quantities**: For quantities involving density derivatives, the NERD model outperforms methods based on Gaussian basis sets. 3. **Versatility of a Single Density Model**: The NERD model not only provides highly accurate density and density derivatives but can also calculate other one-electron observables such as nuclear forces, spin density, and dipole moments through numerical integration. 4. **Scalability and Convergence Rate**: Using the benzene molecule as an example, the scalability and convergence rate of the density model were studied. In summary, the paper proposes a method called NERD, which can extract one-electron density from real-space multi-electron wave functions with high fidelity, demonstrating the potential to obtain highly accurate one-electron density from deep learning wave function representations.

Highly Accurate Real-space Electron Densities with Neural Networks

Interatomic Force from Neural Network Based Variational Quantum Monte Carlo

Ab initio quantum chemistry with neural-network wavefunctions

Deep Neural Network Computes Electron Densities and Energies of a Large Set of Organic Molecules Faster than Density Functional Theory (DFT)

Toward Orbital-Free Density Functional Theory with Small Data Sets and Deep Learning

A real neural network state for quantum chemistry

Density Functionals with Quantum Chemical Accuracy: From Machine Learning to Molecular Dynamics

Accurate Computation of Quantum Excited States with Neural Networks

Low-data deep quantum chemical learning for accurate MP2 and coupled-cluster correlations

Variational Monte Carlo on a Budget -- Fine-tuning pre-trained Neural Wavefunctions

Deep learning of spatial densities in inhomogeneous correlated quantum systems

Deep Density: Circumventing the Kohn-Sham equations via symmetry preserving neural networks

Towards a Foundation Model for Neural Network Wavefunctions

Neural network representation of electronic structure from $ab$ $initio$ molecular dynamics

Improving density matrix electronic structure method by deep learning

Synergy between deep neural networks and the variational Monte Carlo method for small $^4He_N$ clusters

Solving deep-learning density functional theory via variational autoencoders

Deep learning density functional theory Hamiltonian in real space

Accelerating finite-temperature Kohn-Sham density functional theory with deep neural networks

Efficient solutions of fermionic systems using artificial neural networks

Predicting The One-Particle Density Matrix With Machine Learning