Direct Prediction Of Phonon DoS Using Equivariant Neural Network - How To Calculate It And Why Is It Important?
The design, discovery, and property prediction of materials have all benefited greatly from machine learning. However, despite machine learning's effectiveness in predicting discrete qualities, difficulties still exist in predicting continuous properties. Because of the lack of data and crystallographic symmetry issues, the problem is even more difficult in crystalline solids.
Direct prediction of phonon dos using equivariant neural networkwas demonstrated by Zhantao Chen and colleagues. The study was published in Advanced ScienceJournal. Using a limited training samples with over 64 atom types, high-quality prediction is achieved using Euclidean neural networks, which by design are equivariant to 3D rotations, translations, and inversion and so capture entire crystal symmetry.
The predictive model is well-suited to effectively anticipate alloy systems without adding extra computing expense, and it reproduces important properties of experimental data while also generalizing to materials with undiscovered constituents.
Prediction of a diverse set of materials with high phononic specific heat capacity demonstrates the network's potential. This study demonstrates a productive method for probing the phonon structure of materials and may eventually pave the way for speedy screening of candidates for high-efficiency thermal storage and phonon-mediated superconductors.
Phonon is the fundamental excitation in the quantum mechanical description of crystal lattice vibrations. As quanta of excitation of the lattice vibration mode with angular frequency, their energy is k, which is similar to that of photons. Phonons are non-conservable bosons that may be produced or destroyed in collisions.
Because the location of phonons cannot be identified, they are not localized particles. By mixing modes of slightly differing frequency and wavelength, a somewhat confined wavepacket may be produced.
The phonon density of states (DOS) is primarily a function of the local atomic structure and is sensitive to stresses and microstructure at the atomic level.
What is a Phonon? (in English)
With just "first-principles" inputs (atomic species and locations), the team developed a machine learning-based prediction model to directly acquire the high-dimensional material property of phonon DoS. Since Euclidean neural networks are equivariant, they are able to efficiently capture the symmetries of the input crystal. Even with data augmentation, a training set of merely 1200 instances is adequate to make significant predictions.
Elastically strained samples have been linked to the present model, which may be connected to the fact that we limited the possible values of the output energy to the positive integers. However, even in the face of data scarcity, Euclidean neural networks may be used to predict more general features of crystalline solids.
Changing the hot embedding to a weighted multi-hot embedding allows the alloy prediction to have the same computational cost as the pure phase, demonstrating the efficiency of our atomic embedding technique for computing phonon DoS in alloys.
An ML model is data-driven and probabilistic in nature, in contrast to the deterministic nature of ab initio calculations and inelastic scattering for acquiring phonon DoS. As a result, it is not feasible to depend only on a predictive model based on machine learning to get material attributes without doing further validation.
The ML method is powerful, but not only for determining material-specific attribute magnitudes. When used in materials design, ML shows remarkable effectiveness in quickly filtering out alternatives that lack the desired attribute. We find that an entry-level GPU can forecast phonon DoS for all 4346 unknown materials in under 30 minutes.
From the standpoint of property optimization, the ML method looks for and produces a batch of high-performance candidates rather than testing each material individually with high accuracy. For experiments with restricted access to national facilities, the ML model may be helpful in providing direction.
The Dulong-Petit law presents a significant barrier from an applied perspective when looking for potential materials for thermal storage; a highly efficient technique may further facilitate an inverse design from property to suitable structures.
If the density of wave vectors k in the Brillouin zone is uniformly distributed, then the number of phonon modes of a given frequency in a given frequency range may be described by the phonon density of states.
In the quantum mechanical description of vibrations in a crystal lattice, the elementary excitation that is used is called a phonon. Like photons, they excite the lattice in vibrational modes of angular frequency and have an energy of.
The density functional perturbation theory-based phonon database comprising phonon DoS data for roughly 1,500 crystalline materials enables accurate predictions of phonon DoS.
This approach of direct prediction of phonon dos using equivariant neural network offers a potential framework to facilitate high-throughput screening and direct experimental planning for materials with remarkable thermal characteristics. Also, it helps clarify the connections between condensed matter's symmetry, structure, and elementary excitations.