Density of states prediction for materials discovery via contrastive learning from probabilistic embeddings

Abstract

Machine learning for materials discovery has largely focused on predicting an individual scalar rather than multiple related properties, where spectral properties are an important example. Fundamental spectral properties include the phonon density of states (phDOS) and the electronic density of states (eDOS), which individually or collectively are the origins of a breadth of materials observables and functions. Building upon the success of graph attention networks for encoding crystalline materials, we introduce a probabilistic embedding generator specifically tailored to the prediction of spectral properties. Coupled with supervised contrastive learning, our materials-to-spectrum (Mat2Spec) model outperforms state-of-the-art methods for predicting ab initio phDOS and eDOS for crystalline materials. We demonstrate Mat2Spec’s ability to identify eDOS gaps below the Fermi energy, validating predictions with ab initio calculations and thereby discovering candidate thermoelectrics and transparent conductors. Mat2Spec is an exemplar framework for predicting spectral properties of materials via strategically incorporated machine learning techniques.

Examining graph neural networks for crystal structures: limitations and opportunities for capturing periodicity

“…Consequently, machine learning models based on the set of descriptors in this work can approximate 𝑈𝑈 and 𝐶𝐶 v well within the Debye model, which explains why machine learning model based on only descriptors outperforms CGCNN and ALIGNN in Figure 4a, as CGCNN and ALIGNN cannot estimate lattice constants well as in Figure 2b. More discussions about the relation between this work and the prediction of U in Legrain et al 56 and the prediction of phonon density of states by E3NN 27 and Mat2Spec 31 are provided in the Supplementary Information. κ and M are another two properties with improvement from both de-CGCNN and de-ALIGNN (around 10% in these cases).…”

Physically Informed Machine Learning Prediction of Electronic Density of States

“…With regards to model design, we note that there have been significant recent advances in graph neural network designs for materials chemistry applications. − These approaches can be incorporated in our framework with only minor changes to our overall workflow. Additional avenues for improvement can also include better output representations for the DOS, such as using a principal component basis ,, or using autoencoders, , which would help by reducing the output dimensions, as long as the reconstruction errors are sufficiently low. Alternatively, coupling the ML predictions to a physical model such as a tight-binding model or a lower level of DFT theory in a Δ-ML approach could be used to improve accuracy and reduce data requirements.…”

Free energy predictions for crystal stability and synthesisability

“…76 Other efforts using ML for vibrational properties have focused on predicting the full phonon density of states, from which all integrated thermal properties can be derived. 77 Furthermore, one of data sets in the benchmarking test suite, MatBench, 78 As an alternative approach to using calculated vibrational free energies, experimental thermochemical data can also be used for training the ML model. This was done successfully by Bartel et al 79 using the SISSO framework to obtain a closed form analytical expression for the free energy, from which they obtained a test MAE of ∼50 meV/atom over a wide temperature range.…”

“…76 Other efforts using ML for vibrational properties have focused on predicting the full phonon density of states, from which all integrated thermal properties can be derived. 77 Furthermore, one of data sets in the benchmarking test suite, MatBench, 78 is related to phonon dispersions, meaning that we should see active development of ML models capable of predicting thermal properties of materials.…”