Graph Neural Networks for Learning Molecular Excitation Spectra

Abstract

Machine learning (ML) approaches have demonstrated the ability to predict molecular spectra at a fraction of the computational cost of traditional theoretical chemistry methods while maintaining high accuracy. Graph neural networks (GNNs) are particularly promising in this regard, but different types of GNNs have not yet been systematically compared. In this work, we benchmark and analyze five different GNNs for the prediction of excitation spectra from the QM9 dataset of organic molecules. We compare the GNN performance in the obvious runtime measurements, prediction accuracy, and analysis of outliers in the test set. Moreover, through TMAP clustering and statistical analysis, we are able to highlight clear hotspots of high prediction errors as well as optimal spectra prediction for molecules with certain functional groups. This in-depth benchmarking and subsequent analysis protocol lays down a recipe for comparing different ML methods and evaluating dataset quality.

Machine Learning Many-Body Green’s Functions for Molecular Excitation Spectra

“…Data-driven machine learning (ML) has recently been employed to predict density of states (DOS) and quasiparticle energies at the DFT level from only atomic configurations. − Using thousands of (or more) discretized DOS on the real-frequency axis as training data, Gaussian process regression or deep neural network models are trained to predict DOS of organic molecules, bulk crystals, and amorphous materials, although the quality of results often depends on the resolution chosen to smooth the DOS . A different ML approach is based on the SchNet model, where a latent Hamiltonian matrix is first predicted and molecular resonances are obtained as eigenvalues of the effective Hamiltonian. , In the meantime, ML has also been explored to predict GW corrections to quasiparticle energy levels from DFT inputs. − In addition, we note recent works in ML dielectric screening for accelerating GW and Bethe–Salpeter equation (BSE) calculations. , However, to the best of our knowledge, no current ML model predicts photoemission spectra at the general quantum many-body level beyond independent-particle approximation and properly accounts for quasiparticle renormalization and satellites, which are important spectral features in correlated electron systems.…”

Harnessing Deep Neural Networks to Analyze Multi‐Channel Anion Sensing Characteristics of a Ru(II)‐Pyrazolyl‐Bis(Benzimidazole) Complex

“…The technological progress of ML is now manifested in nearly all branches of science and technology [1–11] . Through proper handling of powerful computation and high‐throughput experimentation, ML has expedited the scientific research and technological development [12–31] . Even though the adoption of data‐guided growth of materials is inspiring to recognize the accurate potential of ML models, they should also have the potentiality over solely predictive ability.…”

“…[1][2][3][4][5][6][7][8][9][10][11] Through proper handling of powerful computation and highthroughput experimentation, ML has expedited the scientific research and technological development. [12][13][14][15][16][17][18][19][20][21][22][23][24][25][26][27][28][29][30][31] Even though the adoption of data-guided growth of materials is inspiring to recognize the accurate potential of ML models, they should also have the potentiality over solely predictive ability. The prediction and inner execution of models must offer appropriate explanation to the human specialists.…”

Human brain-inspired chemical artificial intelligence tools for the analysis and prediction of the anion-sensing characteristics of an imidazole-based luminescent Os(ii)-bipyridine complex

“…Obtaining a quantitative insight of a particular observation either through theoretical calculations or by experimentation is always associated with the consumption of time. We employed, in this work, various deep learning and machine learning tools, such as fuzzy logic, [11][12][13][14][15][16] artificial neural network (assisted by three different training algorithms), [17][18][19][20][21][22] adaptive neuro-fuzzy inference system, [23][24][25][26][27] as well as decision tree regression analysis, [28][29][30][31][32][33][34] on our existing anion-sensing dataset of an Os(II)polyheterocyclic complex for proper understanding as well as to fully predict its anion-sensing characteristic within a very short period of time (Chart 1). 35 We utilized herein an imidazolyl bis-benzimidazole-based Os(II) complex having three azole NH motifs in its periphery (Chart 1).…”

“…Recurrent neural network (RNN) and artificial neural network-function fitting (ANN-FF) grids have been employed commonly for this purpose. [17][18][19][20][21][22] In this work, we utilized artificial neural network-function fitting technique because of the static character of the present scheme and also due to its aptitude to recognize and predict complicated systems. Artificial neural network is quite capable of grasping the data but not so skilled with regard to recognizing the significance of each neuron and its weight.…”