Transferable Multilevel Attention Neural Network for Accurate Prediction of Quantum Chemistry Properties via Multitask Learning

The development of efficient models for predicting specific properties through machine learning is great importance the innovation chemistry and material science. However, global electronic structure like Frontier molecular orbital highest occupied (HOMO) lowest unoccupied (LUMO) energy levels their HOMO–LUMO gaps from small-sized molecule data to larger molecules remains a challenge. Here, we develop multilevel attention neural network, named DeepMoleNet, enable chemical interpretable insights being fused into multitask (1) weighting contributions various atoms (2) taking atom-centered symmetry functions (ACSFs) as teacher descriptor. prediction 12 including dipole moment, HOMO, Gibbs free within accuracy achieved by using multiple benchmarks, both at equilibrium nonequilibrium geometries, up 110,000 records in QM9, 400,000 MD17, 280,000 ANI-1ccx random split evaluation. good transferability outside training set demonstrated QM9 Alchemy sets density functional theory (DFT) level. Additional tests on conformations DFT-based MD17 with coupled cluster well public test singlet fission molecules, biomolecules, long oligomers, protein 140 show reasonable predictions thermodynamics properties. proposed network applicable high-throughput screening numerous species spaces accelerate rational designs drug-like candidates, reactions.