Machine learning approach for vibronically renormalized electronic band structures

We present a machine learning (ML) method for efficient computation of vibrational thermal expectation values of physical properties from first principles. Our approach is based on the non-perturbative frozen phonon formulation in which stochastic Monte Carlo algorithm is employed to sample configurations of nuclei in a supercell at finite temperatures based on a first-principles phonon model. A deep-learning neural network is trained to accurately predict physical properties associated with sampled phonon configurations, thus bypassing the time-consuming {\em ab initio} calculations. To incorporate the point-group symmetry of the electronic system into the ML model, group-theoretical methods are used to develop a symmetry-invariant descriptor for phonon configurations in the supercell. We apply our ML approach to compute the temperature dependent electronic energy gap of silicon based on density functional theory (DFT). We show that, with less than a hundred DFT calculations for training the neural network model, an order of magnitude larger number of sampling can be achieved for the computation of the vibrational thermal expectation values. Our work highlights the promising potential of ML techniques for finite temperature first-principles electronic structure methods.

💡 Research Summary

The paper introduces a machine‑learning‑augmented framework for computing temperature‑dependent electronic band structures, focusing on the vibronic (electron‑phonon) renormalization of the band gap. Traditional first‑principles approaches such as density‑functional perturbation theory (DFPT) combined with Allen‑Heine‑Cardona (AHC) theory or GW‑electron‑phonon calculations can accurately capture electron‑phonon coupling, but they are computationally intensive because they require either high‑order perturbative expansions or repeated electronic‑structure evaluations for many phonon configurations. The frozen‑phonon method offers a non‑perturbative alternative: one samples atomic configurations displaced according to the normal‑mode coordinates of the lattice, solves the electronic problem for each configuration, and averages the resulting observables. However, the need to solve an electronic‑structure problem (typically DFT) for each Monte‑Carlo (MC) sample makes the method prohibitive for large supercells or fine temperature grids.

To overcome this bottleneck, the authors propose a two‑stage approach. First, they generate phonon configurations using a harmonic phonon model derived from first‑principles calculations (DFT/DFPT). The normal‑mode coordinates are sampled from Gaussian distributions whose widths are set by the Bose‑Einstein occupation at the target temperature. Crucially, they construct a symmetry‑invariant descriptor for each configuration by applying group‑theoretical techniques: the point‑group symmetry of the crystal is used to transform the raw displacement vectors into a set of generalized coordinates that remain unchanged under symmetry operations. This descriptor dramatically reduces the dimensionality of the input space and guarantees that the neural network respects the underlying crystal symmetry.

Second, a deep neural network (a multilayer perceptron with ReLU activations and Adam optimization) is trained to map the symmetry‑invariant descriptor to the electronic property of interest—in this work, the fundamental band gap. Training data consist of fewer than one hundred DFT calculations, each corresponding to a distinct phonon configuration sampled from the harmonic distribution. The authors employ cross‑validation and early‑stopping to avoid over‑fitting, achieving a model that predicts the band gap with sub‑10 meV error across the sampled configuration space.

The trained model is then used to evaluate thousands of MC‑sampled configurations at various temperatures, effectively replacing the expensive DFT evaluations. The authors demonstrate the method on bulk silicon, a material with a well‑characterized temperature‑dependent band gap. The ML‑augmented MC results reproduce the experimentally observed band‑gap shrinkage (≈ −0.07 eV at 300 K) and remain within 10 meV of reference AHC and GW‑phonon calculations up to 800 K. Compared with a naïve MC‑DFT approach, the new scheme reduces the number of required DFT calculations by roughly an order of magnitude while maintaining comparable accuracy.

Beyond silicon, the authors argue that the framework is broadly applicable. Because the descriptor is independent of the electronic‑structure solver, one can replace DFT with more accurate methods (GW, hybrid functionals, quantum‑chemistry techniques) without altering the ML pipeline. Likewise, the method can be extended to systems with strong electron‑phonon coupling, anharmonic phonons, or low‑dimensional materials, provided a harmonic (or suitably corrected) phonon model is available for sampling. Potential future directions include incorporating anharmonic corrections, simultaneous prediction of multiple temperature‑dependent observables (e.g., dielectric constants, carrier mobilities), and integrating the workflow into high‑throughput materials‑screening platforms.

In summary, the paper delivers three key innovations: (1) a group‑theory‑based, symmetry‑preserving phonon descriptor; (2) a deep‑learning model trained on a modest set of high‑quality DFT data; and (3) a Monte‑Carlo sampling scheme that, when combined with the ML model, yields fast and accurate temperature‑dependent electronic‑structure predictions. This work therefore represents a significant step toward making finite‑temperature first‑principles electronic‑structure calculations practical for large‑scale materials discovery.