OpenNovelty.org

Orbital Transformers for Predicting Wavefunctions in Time-Dependent Density Functional Theory

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
Machine learning density functional theoryTime dependent neural PDE solver
Abstract:

We aim to learn wavefunctions simulated by time-dependent density functional theory (TDDFT), which can be efficiently represented as linear combination coefficients of atomic orbitals. In real-time TDDFT, the electronic wavefunctions of a molecule evolve over time in response to an external excitation, enabling first-principles predictions of physical properties such as optical absorption, electron dynamics, and high-order response. However, conventional real-time TDDFT relies on time-consuming propagation of all occupied states with fine time steps. In this work, we propose OrbEvo, which is based on an equivariant graph transformer architecture and learns to evolve the full electronic wavefunction coefficients across time steps. First, to account for external field, we design an equivariant conditioning to encode both strength and direction of external electric field and break the symmetry from SO(3) to SO(2). Furthermore, we design two OrbEvo models, OrbEvo-WF and OrbEvo-DM, using wavefunction pooling and density matrix as interaction method, respectively. Motivated by the central role of the density functional in TDDFT, OrbEvo-DM encodes the density matrix aggregated from all occupied electronic states into feature vectors via tensor contraction, providing a more intuitive approach to learn the time evolution operator. We adopt a training strategy specifically tailored to limit the error accumulation of time-dependent wavefunctions over autoregressive rollout. To evaluate our approach, we generate TDDFT datasets consisting of 5,000 different molecules in the QM9 dataset and 1,500 molecular configurations of the malonaldehyde molecule in the MD17 dataset. Results show that our OrbEvo model accurately captures quantum dynamics of excited states under external field, including time-dependent wavefunctions, time-dependent dipole moment, and optical absorption spectra characterized by dipole oscillator strength. It also shows strong generalization capability on the diverse molecules in the QM9 dataset.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces OrbEvo, a machine learning model that learns to propagate time-dependent electronic wavefunction coefficients in real-time TDDFT simulations. It resides in the 'Machine Learning and Data-Driven Approaches' leaf of the taxonomy, which contains only four papers total. This is a notably sparse research direction compared to the more established TDDFT and wavefunction-based branches, suggesting that direct ML-based wavefunction evolution remains an emerging and relatively unexplored area within the broader field of time-dependent quantum dynamics.

The taxonomy reveals that OrbEvo's immediate neighbors focus on related but distinct ML strategies: one sibling targets photodynamics trajectories, another addresses excited-state property prediction, and a third explores Hartree-Fock representations. The broader TDDFT branch (containing real-time propagation, linear-response, and hybrid functional methods) represents the traditional computational paradigm that OrbEvo aims to accelerate. By bridging the ML leaf with the real-time TDDFT subcategory, OrbEvo occupies a boundary position—applying data-driven techniques to a problem historically dominated by iterative numerical solvers.

Among 18 candidate papers examined across three contributions, none were flagged as clearly refuting the work. The core OrbEvo model examined 10 candidates with no refutations, the SO(2)-equivariant conditioning examined zero candidates, and the dual interaction methods (wavefunction pooling and density matrix) examined 8 candidates with no refutations. This limited search scope—covering top-K semantic matches and citation expansion—suggests that within the examined literature, no prior work directly overlaps with OrbEvo's specific combination of equivariant graph transformers, external field conditioning, and dual wavefunction-density matrix interaction schemes for TDDFT propagation.

Given the sparse ML-for-TDDFT landscape and the absence of refutations among 18 examined candidates, the work appears to occupy a relatively novel niche. However, the small search scale means the analysis cannot rule out relevant prior work outside the top-K semantic neighborhood or in adjacent subfields. The taxonomy structure indicates that while ML approaches to quantum dynamics are growing, direct wavefunction evolution via transformers remains less crowded than traditional TDDFT or nonadiabatic methods.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
18
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: predicting time-dependent wavefunctions in molecular systems. This field encompasses a diverse set of computational strategies for capturing how electronic and nuclear degrees of freedom evolve under external perturbations or during photochemical events. The taxonomy reflects several major branches: Time-Dependent Density Functional Theory Methods (TDDFT) offer a balance between accuracy and computational cost for many-electron systems, with foundational implementations like TDDFT Response Implementation[12] and ongoing refinements such as Symmetric Quasiclassical TDDFT[1] and Low-Rank Hybrid TDDFT[14]. Wavefunction-Based Quantum Dynamics Methods provide high-accuracy treatments through approaches like Multiconfigurational MCTDH[16] and Time-Dependent DMRG[18], while Nonadiabatic and Mixed Quantum-Classical Dynamics address coupled electronic-nuclear motion via methods exemplified by Direct Nonadiabatic Dynamics[33] and Exact Factorization Analysis[8]. Machine Learning and Data-Driven Approaches represent an emerging direction that leverages neural architectures to accelerate or replace traditional quantum propagation. On-the-Fly and Adaptive Potential Energy Surface Methods enable dynamics without precomputed surfaces, and specialized branches focus on Photochemistry and Excited State Applications as well as Multiscale Modeling for complex environments. Within the Machine Learning and Data-Driven Approaches branch, a small handful of works explore how neural networks can learn quantum dynamics directly from data. Machine Learning Photodynamics[5] and Machine Learning Excited States[29] demonstrate early efforts to predict excited-state properties and nonadiabatic trajectories using data-driven models, while Scalable Hartree-Fock Learning[30] targets efficient representations of electronic structure. Orbital Transformers[0] sits naturally in this cluster, employing transformer architectures to predict time-evolved molecular orbitals and wavefunctions, thereby bypassing expensive iterative solvers. Compared to Machine Learning Photodynamics[5], which focuses on trajectory-level predictions, Orbital Transformers[0] emphasizes direct wavefunction propagation at the orbital level. This distinction highlights a key trade-off in the field: whether to learn coarse-grained dynamics or fine-grained quantum amplitudes, with the former offering broader applicability and the latter promising higher fidelity for systems where explicit wavefunction detail is essential.

Claimed Contributions

OrbEvo model for learning time-dependent wavefunction evolution in TDDFT

The authors introduce OrbEvo, a machine learning model that uses an equivariant graph transformer to predict the temporal evolution of electronic wavefunction coefficients in real-time time-dependent density functional theory. The model learns to propagate wavefunctions under external electric fields, enabling efficient prediction of quantum dynamics.

10 retrieved papers
SO(2)-equivariant conditioning for external electric field

The authors develop a method to incorporate external electric field information into the model by breaking the full SO(3) rotational symmetry down to SO(2) symmetry around the field axis. This conditioning encodes both the magnitude and direction of the field while respecting the reduced symmetry constraints.

0 retrieved papers
Two interaction methods for electronic states: wavefunction pooling and density matrix

The authors propose two distinct architectures for handling interactions among multiple electronic states. OrbEvo-WF uses layer-wise pooling over electronic states, while OrbEvo-DM aggregates information via density matrix features computed through tensor contraction, providing a more physically motivated approach aligned with TDDFT formalism.

8 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

OrbEvo model for learning time-dependent wavefunction evolution in TDDFT

The authors introduce OrbEvo, a machine learning model that uses an equivariant graph transformer to predict the temporal evolution of electronic wavefunction coefficients in real-time time-dependent density functional theory. The model learns to propagate wavefunctions under external electric fields, enabling efficient prediction of quantum dynamics.

Contribution

SO(2)-equivariant conditioning for external electric field

The authors develop a method to incorporate external electric field information into the model by breaking the full SO(3) rotational symmetry down to SO(2) symmetry around the field axis. This conditioning encodes both the magnitude and direction of the field while respecting the reduced symmetry constraints.

Contribution

Two interaction methods for electronic states: wavefunction pooling and density matrix

The authors propose two distinct architectures for handling interactions among multiple electronic states. OrbEvo-WF uses layer-wise pooling over electronic states, while OrbEvo-DM aggregates information via density matrix features computed through tensor contraction, providing a more physically motivated approach aligned with TDDFT formalism.

Table of Contents

Orbital Transformers for Predicting Wavefunctions in Time-Dependent Density Functional Theory | Novelty Validation