Machine learning the local electronic density of states
A. Aryanpour, Ali Sadeghi · 2025
Read Machine learning the local electronic density of states by A. Aryanpour, Ali Sadeghi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-16052.txt.
Introduction
A machine learning study by Aryanpour and Ali Sadeghi, published in 2025, teaches models the local density of electronic states atom by atom. That measure shapes many traits of a material, from energy to heat response. Atom level learning promises maps that scale and transfer across structures.
The model learns each atom's share of the state spectrum rather than a whole structure score. It hits high accuracy on pure and mixed silicon and carbon forms of many sizes. A harder tin sulfur selenium test shows the approach travels to complex compounds. Local predictions then lift forecasts of band energy, Fermi energy, heat capacity, and magnetic response. The route skips costly first principles math for fresh structures.
Catalog record
| CREATOR | A. Aryanpour, Ali Sadeghi |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 44 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-16052
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Machine learning the local electronic density of states A. Aryanpour, Ali Sadeghi Abstract Electronic density of states (DOS) plays a crucial role in determining and understanding materials properties. We investigate the machine learnability of additive atomic contributions to electronic DOS, focusing on atom-projected DOS rather than structural DOS. This approach for structure-property mapping is both scalable and transferable, and achieves high prediction accuracy for pure and compound silicon and carbon structures of various sizes and configurations. Furthermore, we demonstrate the generalizability of this model to complex Sn-S-Se compound structures. Utilizing locally trained DOS is shown to significantly enhance the accuracy of predicting material properties, including band energy, Fermi energy, heat capacity, and magnetic susceptibility. Our findings indicate that directly learning atomic DOS, rather than structural DOS, improves the efficiency, accuracy, and interpretability of machine learning in structure-property mapping. This streamlined approach reduces computational complexity, paving the way for examination of electronic structures in materials without the need for computationally expensive ab initio calculations. I Introduction In the last decade, low-cost calculation of the electronic structure of materials has increasingly been taken into consideration for both pure research and applied purposes. Compared to quantum chemistry methods which determine the many-body wavefunction, the density functional theory (DFT) provides a much faster computational scheme by solving the single-electron Kohn-Sham equations for the ground state electron density [ 1 , 2 ] . Although the reliability of DFT calculations has been continuously improved through the decades [ 3 ] , its application is still limited to systems that are only a few orders of magnitude larger than those tractable with quantum chemistry methods. The computational bottleneck of DFT lies in the self-consistent construction of large Hamiltonian matrices, which involves repeated diagonalization. The unknown true form of the exchange and correlation (XC) functionals also limits the accuracy of this approach. To speed up electronic structure calculations and extend its application to larger systems, one may approximate the DFT Hamiltonian with a tight binding-type Hamiltonian which can partially be filled with pre-calculated reference values [ 4 ] or trained on ab initio eigenvalues by machine learning (ML) [ 5 , 6 ] . Alternatively, an ultimate outcome of electronic structure calculations, either scalar, vectorial or tensorial quantities, might be directly predicted with ML. For example, the potential energy surface of an atomic structure or the Hellmann-Feynman forces on the atomic centers as required for molecular dynamics simulations can be the output layer of an artificial neural network. [ 7 , 8 , 9 ] This approach bypasses the costly construction and digitalization of the Hamiltonian matrices but it also ignores the whole physics of the problem. The distribution of electronic energy levels of a system, known as the density of states (DOS), is a key quantity in analyzing physical phenomena and designing materials with desired electronic and optical properties. Although the structural DOS is a global quantity that depends on the atomic structure as a whole, it is a common practice to decompose it into orbital- projected components, known as PDOS. Then one can interpret physical properties and chemical behaviour in terms of contributions from the constituting atoms or atomic orbitals. Such an understanding is helpful in designing catalysts, semiconductor devices, or nanostructure materials, where electronic characteristics can substantially vary at the nanoscale. This is also crucial when the functionality of the material critically depends on heterogeneity at the atomic scale. Generally, decomposition of a physical quantity into atomic contributions leads to two important computational advantages: scalability and transferability. The atomic contributions are first trained, with low computational effort, on small-sized and basic structures but are then applicable to systems of much larger sizes and complexities. The underlying assumption is locality of atomic contribution. The intuition behind locality is mostly based on the nearsightedness principle , referring to the pioneering work by Kohn and Prodan [ 2 , 10 , *Prodan06 ] who showed that for a many-electron system the electron density at a given point is practically insensitive to the change of the external potential at large distances. In practice, it has also been shown by DFT calculations that in the absence of chemically impactful defects like aliovalent impurities or vacancies, the electronic matter in bulk (but not in low dimensional samples) remains fairly nearsighted. [ 12 ] We aim to explore in this study whether the atom-projected DOS, referred to as local DOS (LDOS) hereafter, depends merely on the local environment of the atoms. DFT calculations provide DOS (either structural DOS or atomic LDOS as the sum of PDOS over orbitals of a given atom) as an array of real numbers distributed on the energy interval. The array might serve as the input feature vector into a ML model for predicting physical quantities [ 13 , 14 ] , topological invariants [ 15 ] and so on. Alternatively, DOS, as an array, can be the target of ML prediction. [ 16 , 17 ] In this study we aim to show the highly employability of the locality nature of atomic LDOS for a supervised learning and thus an improved efficiency of learning LDOS instead of structural DOS. In the rest of this manuscript, we first explain the mathematical background behind local learning of DOS and then introduce our ML procedure and training dataset. The prediction error for LDOS and DOS are compared for selected silicon and carbon systems. We then show the performance of predication of a few of the derived physical properties from the trained LDOS, including band energy, Fermi energy and DOS at the Fermi level, magnetic susceptibility and excitation distribution spectrum before we draw our conclusions. II Theoretical Background and Methods II.1 Local Density of States We denote the total (or structural ) DOS by \mathcal{D} , such that \mathcal{D}(\varepsilon)\,d\varepsilon represents the number of states available for electrons within the energy interval \varepsilon to \varepsilon+d\varepsilon throughout the entire structure. The central idea of this work is that the structural DOS can be split up as \mathcal{D}(\varepsilon)=\sum_{i}\mathcal{D}_{i}(\varepsilon), (1) to atomic contributions \mathcal{D}_{i}(\varepsilon) which can be defined in several ways [ 18 ] . We emphasize that the method used in Ref. 16 is not one of the physical projection schemes explained in the following. Instead, they machine learned \mathcal{D}_{i} by minimizing the difference between the total DOS serving as the reference and the predicted sum of atomic \mathcal{D}_{i} , as subsequently discussed. When necessary, we differentiate between physical projection and machine learning splitting of DOS by referring to them as projected DOS and split DOS, respectively. For a crystalline system with a Brillouin zone (BZ) of volume \Omega_{\text{BZ}} , the DOS is given by \mathcal{D}(\varepsilon)=\frac{1}{\Omega_{\text{BZ}}}\sum_{n}\int_{\text{BZ}}\delta\left(\varepsilon-\varepsilon_{n}(\mathbf{k})\right)\,d\mathbf{k}. (2) Electronic structure calculations provide eigenfunctions |\psi_{n}(\mathbf{k})\rangle and corresponding energies \varepsilon_{n}(\mathbf{k}) at a finite number N_{k} of grid points in {\mathbf{k}} -space. Thus, the latter equation is approximated by \mathcal{D}(\varepsilon)=\frac{1}{N_{\mathbf{k}}}\sum_{n,{\mathbf{k}}}\delta(\varepsilon-\varepsilon_{n,\mathbf{k}})\langle\psi_{n\mathbf{k}}|\left(\int|{\bf r}\rangle\langle{\bf r}|\,d{\bf r}\right)|\psi_{n\mathbf{k}}\rangle\\ =\frac{1}{N_{\mathbf{k}}}\sum_{n,{\mathbf{k}}}\int\left|\psi_{n\mathbf{k}}({\bf r})\right|^{2}\delta(\varepsilon-\varepsilon_{n,\mathbf{k}})\,d{\bf r}. (3) Note that we have employed the normalization of the eigenfunctions 1=\langle\psi_{n\mathbf{k}}|\psi_{n\mathbf{k}}\rangle and the completeness relation 1=\int|{\bf r}\rangle\langle{\bf r}|\,d{\bf r} . For a non-periodic sample such as a molecule or atomic cluster, N_{\mathbf{k}}=1 , and calculations are performed only at the point {\mathbf{k}}=(0,0,0) . To transform this equation into Eq. ( 1 ), the whole space is partitioned into non-overlapping atomic