The optical properties of defects within solid materials support a wide range of phenomena, ranging from the vibrant colors of gemstones to the potential for single photon emissions of quantum networks. To accurately model these optical transitions, we need to understand how electrons interact with atomic oscillations called phonons, but calculating these interactions traditionally requires immeasurable computational resources. Mark E. Tulianski, John L. Lyons, and Norm Bernstein all demonstrate how to bypass this limitation by using machine learning to predict the phonon spectrum, all at the U.S. Navy Institute. This innovative approach greatly accelerates calculations without sacrificing accuracy, opens a new path for a detailed investigation of defect vibrational properties, and allows for the resolution of subtle features such as complex vibrational modes that affect the emission of silicon defects.
However, the current method requires computationally expensive evaluation of all phonon modes in simulation cells containing hundreds of atoms. This work demonstrates that this bottleneck can be overcome using inter-machine learning potentials with negligible accuracy losses. An important finding is that atomic relaxation data from regular first-principles calculations are sufficient as a data set for fine tuning, but additional data could further improve model performance. The efficiency of this approach allows for the study of defect vibrational properties with high levels of theory, and the team fine-tunes the model to hybrid functional calculations to obtain highly accurate spectra and compares the results with explicit calculations and experiments of various defects.
Defect formation energy via density functional theory
This supplementary material provides a detailed explanation of the calculation methods used to calculate defect properties of materials using density functional theory (DFT). The authors carefully document the approach, provide transparency and reproducibility, and specify implementations of DFT using a VASP software package that includes exchange correlation capabilities and selection of variance corrections. They detail the parameters used for K-point sampling, energy convergence, and force convergence during structural relaxation, model defects using supercell methods, and utilize doped packages to facilitate creation. The authors address the complexity of charge state ambiguity in defect calculations and explain how to correct it, and perform excited state calculations using time-dependent DFTs.
They describe the use of mean points in Brillouin zones for defect calculations, provide an overview of the DFTs and exchange correlation features used, and discuss the Jahn-Teller effects and bibronic structures related to understanding defect properties. The divergence of Kullback-Leibler is mentioned as a measure of differences in probability distributions, and the performance of the scan meta GGA feature is evaluated for defect calculations. The author carefully considers the size of the supercells, shows convergence on the K-point grid, and compares the results obtained with various features, including PBE, HSE06, and scan. They explain how they determined the most stable charge state of the defect, discussed the bibronic coupling associated with optical properties, and how the optical absorption spectra were calculated. The team's approach utilizes machine learning to model complex interactions between electrons and atomic oscillations. This is a traditionally computationally intensive process that requires extensive calculations. This new approach dramatically reduces computational burden and allows for detailed studies of previously impractical defects. The core of this innovation lies in the use of “basic models,” pre-trained with existing data, machine learning algorithms, refined using surprisingly small data sets unique to each material defect.
Initial tests show that the basic model can qualitatively predict luminescence, but further improvements are important for accurate results. Surprisingly, the team found that data from routine calculations already being performed to determine the equilibrium geometry of defects, is sufficient to significantly improve predictive power and to effectively provide a “free” data set. To further improve accuracy, researchers investigated generating a small number of additional configurations and found that adding just 10 new data points significantly improves the performance of the model. For the carbon defects of hexagonal boron nitride, this approach provided nearly 150 times faster speed compared to traditional methods, while maintaining high accuracy. The team has committed to applying this sophisticated method to several complex materials, achieving detailed calculations and quantitative agreements, opening doors for direct comparisons with experimental data, and accelerating the discovery and design of materials with customized optical properties.
Machine learning streamlines vibration analysis of defects
This study presents a new approach to calculating the vibrational properties of defects in solid materials. This is important for understanding optical behavior and potential applications in areas such as quantum computing. The team applied cross-machine learning potential (MLIP) to predict these properties, achieving accuracy comparable to traditional computational methods. By training MLIPs on data from routine calculations of atomic relaxation, they significantly reduced the computational burden and allowed for the study of much larger systems, including 8000 atomic ratings of defects in silicon than previously feasible. A key advance is streamlining the process of calculating how defects interact with vibrations, which directly affects the emission spectrum. This work establishes systematic methods for developing and improving MLIP specifically for defect research, opening the door to applying these technologies to a wider range of materials and defects. Although the current study utilized specific hybrid functions within density functional theory, the authors emphasize the adaptability of approaches to other theoretical levels, acknowledging that further improvements in MLIP enhance the ability to accurately analyze and characterize defects from Indigenous principles.
