Application of artificial neural networks and genetic algorithms to modeling molecular electronic spectra in solution

Abstract

Improving the accuracy of density-functional theory calculation: The genetic algorithm and neural network approach A novel approach is presented for finding the vibrational frequencies, Franck-Condon factors, and vibronic linewidths that best reproduce typical, poorly resolved electronic absorption ͑or fluorescence͒ spectra of molecules in condensed phases. While calculation of the theoretical spectrum from the molecular parameters is straightforward within the harmonic oscillator approximation for the vibrations, ''inversion'' of an experimental spectrum to deduce these parameters is not. Standard nonlinear least-squares fitting methods such as Levenberg-Marquardt are highly susceptible to becoming trapped in local minima in the error function unless very good initial guesses for the molecular parameters are made. Here we employ a genetic algorithm to force a broad search through parameter space and couple it with the Levenberg-Marquardt method to speed convergence to each local minimum. In addition, a neural network trained on a large set of synthetic spectra is used to provide an initial guess for the fitting parameters and to narrow the range searched by the genetic algorithm. The combined algorithm provides excellent fits to a variety of single-mode absorption spectra with experimentally negligible errors in the parameters. It converges more rapidly than the genetic algorithm alone and more reliably than the Levenberg-Marquardt method alone, and is robust in the presence of spectral noise. Extensions to multimode systems, and/or to include other spectroscopic data such as resonance Raman intensities, are straightforward.

Optimization and validation of a deep learning CuZr atomistic potential: Robust applications for crystalline and amorphous phases with near-DFT accuracy

“…On the other hand, ML potentials do not employ any explicit functional form for the dependence of the energies and forces on the atomic coordinates, but rather "learn" how atoms interact from a statistical model that relies on a massive dataset typically obtained from DFT calculations. [24][25][26][27][28] Similar to the traditional counterparts, ML potentials also suffer from transferability errors associated with atomic environments that are not included in the training. In fact, the transferability of ML potentials could be even worse than standard potentials given the lack of any physical intuition in the model.…”

Convergence acceleration in machine learning potentials for atomistic simulations

“…[13][14][15][16][17][18][19] Machine learning potentials (MLPs) allow simulating atomistic interaction energies and forces without any explicit functional form, distinct from the specific functional interactions of conventional potentials. [20][21][22][23][24] However, this freedom comes at the cost of requiring a large dataset typically generated by DFT for training. For instance, the Behler group developed a Cu MLP using 35k configurations 25 and a model for CuZnO using 100k configurations 26 .…”

Automated parameter optimization in modeling absorption spectra and resonance Raman excitation profiles

“…Motivated by the observation that an experienced human operator can often make very good initial ''guesses'' for the fitting parameters, we trained a neural network to return the initial estimates for the vibronic parameters used to generate optical absorption spectra, and then coupled this with a genetic algorithm to refine the parameters and a final Levenberg-Marquardt step for final refinement. 22 While some success was achieved, the combination of three methods resulted in a complicated algorithm and the very poor scaling of the neural network training time with size of parameter space limited the method to small molecules having only a few vibrational modes. Raman intensity calculations were not attempted with this method.…”