Research map / Probabilistic and scientific modeling / Molecules and proteins
Gaussian Plane-Wave Neural Operator for Electron Density Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GPWNO, a neural operator for electron density estimation that sums a plane-wave (PW) neural operator branch on a lattice-based probe grid with an atom-centered Gaussian-type orbital (GTO) branch, with high-frequency masking, outperforming ten baselines on QM9, MD, and Materials Project datasets.
Problem
Predicting electron density of molecules and crystals accurately and cheaply, since DFT is cubic in system size, and plane-wave neural operators struggle to represent high-frequency density near atoms with a realistic basis size.
Approach
Decompose the density into a PW-basis neural operator (low-frequency, periodic, lattice-based probe nodes with FFT convolution) plus an atom-wise GTO branch (equivariant TFN-style message passing) and mask the PW branch near atoms so the two bases complement each other.
Key results
- On QM9 (unrotated), GPWNO reaches 0.73% NMAE versus 0.93% for the second-best InfGCN.
- On MD, GPWNO reaches 2.45% NMAE on benzene and 4.00% on ethanol, compared with 5.11% and 8.43% for InfGCN.
- On the Materials Project mixed set, GPWNO reaches 4.84% NMAE versus 5.35% for InfGCN, and it is best in all seven crystal families.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GPWNO instead of InfGCN | Electron density estimation · QM9 | NMAE (%): 0.93 -> 0.73 | Helped |
| Use GPWNO instead of InfGCN | Electron density estimation · MD molecules | NMAE (%): 8.43 -> 4.00; 5.11 -> 2.45; 5.51 -> 2.68; 5.95 -> 2.73; 7.01 -> 3.67; 10.34 -> 5.32 | Helped |
| Use GPWNO instead of InfGCN | Electron density estimation · MP crystal families | NMAE (%): 5.35 -> 4.84; 4.36 -> 3.92; 4.63 -> 4.44; 5.06 -> 4.73; 5.02 -> 4.52; 5.21 -> 4.92; 4.96 -> 4.55; 4.93 -> 4.32 | Helped |
| Add GTO and HFM to PW-only prediction (ablation) | GPWNO · Benzene | NMAE (%): 6.92 -> 2.45 | Helped |
| Add PW and HFM to GTO-only prediction (ablation) | GPWNO · Benzene | NMAE (%): 2.90 -> 2.45 | Helped |
| Add HFM to PW + GTO (ablation) | GPWNO · Benzene | NMAE (%): 2.62 -> 2.45 | Helped |
| Add GTO and HFM to PW-only prediction (ablation) | GPWNO · MP Mixed | NMAE (%): 48.69 -> 4.84 | Helped |
| Add PW and HFM to GTO-only prediction (ablation) | GPWNO · MP Mixed | NMAE (%): 5.35 -> 4.84 | Helped |
| Add HFM to PW + GTO (ablation) | GPWNO · MP Mixed | NMAE (%): 5.00 -> 4.84 | Helped |
Surprising findings
- A plane-wave-only model performs very poorly (e.g., 48.69% NMAE on the MP mixed set) without the GTO branch, indicating PW alone cannot capture high-frequency density.
- FNO performs poorly despite its Fourier-based similarity to the PW approach, which the authors attribute to inefficient memory usage limiting high-frequency expressiveness.
Limitations
- Evaluation uses NMAE on density only; downstream quantities such as exchange-correlation energy are not evaluated.
- High-frequency masking introduces discontinuities near the cutoff boundary, which the authors deem ignorable empirically.
- Baseline results are partly taken from prior papers or re-run with official code, and some datasets (e.g., Cubic2) are reported only in the appendix.
Experimental setup
Datasets: QM9 MD (ethanol, benzene, phenol, resorcinol, ethane, malonaldehyde) Materials Project (seven crystal families and mixed) Cubic2
Benchmarks: QM9 MD Materials Project crystal families Cubic2
Metrics: NMAE
Baselines: InfGCN CNN DeepDFT DeepDFT2 EGNN DimeNet DimeNet++ GNO FNO LNO
Builds on: Fourier neural operator InfGCN tensor field network DeepDFT
Closest papers by meaning
- A Recipe for Charge Density Prediction (NeurIPS 2024 · similarity 0.89)
- Learning the Electronic Hamiltonian of Large Atomic Structures (ICML 2025 · similarity 0.87)
- Neural Pfaffians: Solving Many Many-Electron Schrödinger Equations (NeurIPS 2024 · similarity 0.87)
- Enhancing the Scalability and Applicability of Kohn-Sham Hamiltonians for Molecular Systems (ICLR 2025 · similarity 0.87)
- Learning Generalized Residual Exchange-Correlation-Uncertain Functional for Density Functional Theory (AAAI 2025 · similarity 0.86)
- EDBench: Large-Scale Electron Density Data for Molecular Modeling (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.