Neural networks trained on molecule pairs predict light absorption in stacks of fifty

The work

Technical · from the record

Neural networks were trained on quantum-chemistry reference data for molecular dimers to predict the elements of a Frenkel exciton Hamiltonian — local-excitation and charge-transfer state energies and their couplings — from dimer geometry alone. Because the authors show a homogeneous aggregate's Hamiltonian can be assembled from its constituent dimer pairs, the dimer-trained model was used to build Hamiltonians for trimers, tetramers and perylene aggregates of up to 50 monomers, and analytic approximations of the coupling terms were introduced to fix the sign (phase) of the off-diagonal elements. Against out-of-sample perylene and tetracene test sets of 500 conformations each, predicted excitation energies differed from the ab initio exciton model by a mean absolute error of around 15 meV and from all-electron TDDFT by around 30 meV for adiabatic states below the number of monomers, with larger errors for higher states. Diagonalising the predicted Hamiltonians gave an optical gap decreasing by 0.48 eV from the isolated monomer to an infinite-sized aggregate, with predicted gaps of 3.18 eV for the monomer and 2.70 eV for the infinite aggregate against experimental values of 2.98 eV and 2.58 eV.

How AI was used

Classical molecular dynamics of a condensed perylene system supplied dimer conformations, which were filtered by geometric criteria and thinned with furthest-point sampling into three 8,000-member subsets (face-to-face, head-to-tail or T-shaped, and separated) forming the 24,000-dimer PrDim training set, with an equivalent TtDim set for tetracene. Reference Frenkel Hamiltonian elements were computed with TeraChem's ab initio exciton model at the wB97X-D3/6-31G* level for two subsets and by stated approximations reusing those results for the separated subset. A modified TorchANI architecture was then trained to map dimer coordinates to Hamiltonian elements: networks were specialised by matrix-element type rather than chemical element, and each atom's environment vector was augmented with a Boolean monomer-belonging label and with the atomic decomposition of an analytic approximation of the target element, obtained by aligning a DFT-optimised reference monomer wave function with RESP and TrESP charges onto each monomer and computing Coulomb interactions or frontier-orbital overlap integrals. Separate models were trained for perylene and tetracene on an 80/20 random train/test split with fixed hyperparameters, PyTorch, 500 epochs, batch size 64 and an Adam optimiser with an exponentially decaying learning rate, minimising summed mean squared error over matrix elements. For larger assemblies the dimer-level predictions were assembled into the aggregate Hamiltonian, with charge-transfer/charge-transfer couplings that do not appear in dimers approximated from predicted local-excitation/charge-transfer couplings, and the unmodified approximation signs used to set the coupling phases; diagonalisation then gave excitation energies and eigenvectors, and aligned reference transition dipole moments weighted by the eigenvector coefficients gave S1 oscillator strengths.