2025 · Physica Scripta
Enabling ease of access to quantum chemistry with transformer-based text encoding and physics-informed multilayer perceptron
Can Polat , Hasan Kurban , Mustafa KurbanResearch highlights
Text-To-Eigenvalue: a text interface to property prediction
Combine encoded material attributes with physical information for quantum-property prediction.
Why we investigated it
Quantum simulations require computational resources and specialist input preparation, motivating more accessible surrogate representations.
What the work contributes
T2E feeds text-derived vectors and physical attributes into a multilayer perceptron.
What the evidence shows
The study evaluates elements, small molecules, and molecular benchmarks including QM9 and MD17 against neural prediction baselines.
Build on this work
Use the paper to explore how textual and physical descriptors can be combined for a defined property-prediction task.
- Text-To-Eigenvalue
- Quantum chemistry
- Text encoding
- Property prediction
Abstract
Abstract Accurate prediction of large-scale material properties is a critical challenge in materials science, driven by the need for efficient computational methods to explore and design advanced materials for various applications. Traditional approaches often rely on quantum chemistry simulations, which are computationally intensive and require domain-specific expertise. This work introduces Text-To-Eigenvalue ( T2E ), an innovative approach that combines text-to-vector encoding with a multilayer perceptron for rapid and precise material simulations. T2E converts key material attributes into vector representations, processed through an MLP framework enriched with critical physical data. This fusion of textual, physical, and quantum information enables T2E to predict material properties efficiently and accurately, significantly reducing the computational burden typically associated with particle interactions and eliminating the need for extensive quantum chemistry expertise. Rigorous validation across diverse elements and small molecules demonstrates that T2E outperforms state-of-the-art (SOTA) models like FermiNet and PsiFormer . Moreover, T2E successfully predicts multiple physical properties using the benchmark molecule datasets, such as QM9 and MD17, underscoring its versatility and effectiveness when compared SOTA graph neural networks.
Cite
@article{polat2025enabling,
title = {Enabling ease of access to quantum chemistry with transformer-based text encoding and physics-informed multilayer perceptron},
author = {Can Polat and Hasan Kurban and Mustafa Kurban},
year = 2025,
journal = {Physica Scripta},
doi = {10.1088/1402-4896/adf14b},
}