Nikita Fedik1,Maksim Kulichenko1,Nicholas Lubbers1,Kipton Barros1,Benjamin Nebgen1,Sergei Tretiak1
Los Alamos National Laboratory1
Nikita Fedik1,Maksim Kulichenko1,Nicholas Lubbers1,Kipton Barros1,Benjamin Nebgen1,Sergei Tretiak1
Los Alamos National Laboratory1
Recent advancements in atomistic machine learning model have showcased significant progress in chemistry and material science. However, these approaches often face challenges when applied to unexplored regions of chemical space. Interatomic potential models, for instance, lack crucial electronic structure information, often leading to limited transferability. To address these limitations, our research group has been dedicated to developing PYSEQM (Pytorch-based Semiempirical Quantum Mechanics), a differentiable physics model that combines domain knowledge of semiempirical quantum mechanics with machine learning as a corrective tool. Paired with atomistic neural network backend, PYSEQM allows backpropagation through Hamiltonian, replacing atom-type dependent constants with structure-aware parameters generated on-the-fly.<br/>This presentation will span the release of PYSEQM2.0 which expands capabilities to excited states through iterative solutions in Krylov subspace and dynamics beyond the ground state. GPU-based batched Davidson algorithm extends simulations to large organic molecules relevant to photovoltaics, organic solar cells and photoinduced processes.