Machine learning (ML) is increasingly used in chemical physics and materials science. One major area of thrust is machine learning of properties of molecules and solid materials from descriptors of chemical composition and structure. Recently, kernel regression methods of various flavors—such as kernel ridge regression, Gaussian process regression, and support vector machine—have attracted attention in such applications. Kernel methods allow benefiting simultaneously from the advantages of linear regressions and the superior expressive power of nonlinear kernels. In many applications, kernel methods are used in high-dimensional feature spaces, where sampling with training data is bound to be sparse and where effects specific to high-dimensional spaces significantly affect the performance of the method. We review recent applications of kernel-based methods for the prediction of properties of molecules and materials from descriptors of chemical composition and structure and related purposes. We discuss methodological aspects including choices of kernels appropriate for different applications, effects of dimensionality, and ways to balance expressive power and reliability of the model in high-dimensional feature spaces and with sparse data. We also discuss kernel regression-based hybrid ML approaches.
Feed-forward neural networks (NNs) are widely used for the machine learning of properties of materials and molecules from descriptors of their composition and structure (materials informatics) as well as in other physics and chemistry applications. Often, multilayer (so-called “deep”) NNs are used. Considering that universal approximator properties hold for single-hidden-layer NNs, we compare here the performance of single-hidden-layer NNs (SLNN) with that of multilayer NNs (MLNN), including those previously reported in different applications. We consider three representative cases: the prediction of the band gaps of two-dimensional materials, prediction of the reorganization energies of oligomers, and prediction of the formation energies of polyaromatic hydrocarbons. In all cases, results as good as or better than those obtained with an MLNN could be obtained with an SLNN, and with a much smaller number of neurons. As SLNNs offer a number of advantages (including ease of construction and use, more favorable scaling of the number of nonlinear parameters, and ease of the modulation of properties of the NN model by the choice of the neuron activation function), we hope that this work will entice researchers to have a closer look at when an MLNN is genuinely needed and when an SLNN could be sufficient.
Ab initio modeling of plasticity has been a long-standing challenge. Because of the intrinsically large-scale nature of phenomena such as plasticity and fracture, such calculations are typically done with force fields that are simplistic approximations of complex interatomic interactions and cannot provide insight into electronic structure-mediated mechanisms, which is required especially in the presence of chemical changes, for example in embrittlement. Ab initio-based modeling is desired for better descriptive and predictive power, to benchmark force fields at the scale of the phenomena, and to gain mechanistic understanding necessary for rational rather than ad hoc material design. We overview the status and perspectives of ab initio-based modeling of plasticity-related phenomena. We first consider small-scale modeling with DFT (density functional theory) that is often used to infer plastic properties, before focusing on direct larger-scale ab initio-based modeling of such phenomena with methods including density functional tight binding, orbital-free (OF) DFT, and DFT/OF-DFT coupled with molecular mechanics (QM/MM) or quasicontinuum approaches that are able to treat billions of atoms. We highlight elements of insight gained in respective works which were only possible at the electronic structure level. We conclude that with the ongoing progress in large-scale electronic structure methods—both those that already have been used for simulations of plasticity and those that still have not but carry promise in this domain—direct ab initio level modeling of plasticity and related phenomena will become more important in the near future, increasing the impact of atomistic modeling on materials science.