BEGIN:VCALENDAR VERSION:2.0 PRODID:-//ZContent.net//ZapCalLib 1.0//EN CALSCALE:GREGORIAN METHOD:PUBLISH BEGIN:VEVENT SUMMARY:Emergent-Scale Prediction of Molecular Properties with Chemically-I nformed Neural Networks LOCATION:Chemistry A101 TZID:America/Denver DTSTART:20260507T160000 UID:2026-05-31-02-04-48@natsci.colostate.edu DTSTAMP:20260531T020448 Description:Seminar Abstract:\n\nMachine learning (ML) models have emerged as powerful tools for scientific discovery. Leveraging modern computing ha rdware allows for the generation of large data sets which can be used to t rain accurate\, generalizable models to predict the physico-chemical behav ior of molecular systems. However\, the predictive power of such models is bounded by how chemical knowledge is encoded within the model. Unlike gen eral-purpose neural networks\, which treat molecular data as abstract nume rical arrays\, chemically-informed architectures incorporate physical and structural constraints directly into their mathematical form. This talk ex amines the design principles behind such architectures and traces their ap plication across different intra- and intermolecular scales of organizatio n.\n\nAt the level of individual molecules\, graph neural networks represe nt atoms as nodes and bonds as edges\, propagating information through the molecular graph in a manner that mirrors molecular topology. We will exam ine the results of different parameterizations of this information in the context of predicting bond dissociation enthalpies and adiabatic singlet-t riplet excitation energies\, and present novel analytic tools which connec t learned model behaviors with chemical intuition.\n\nWe then extend this approach to predict the interaction behaviors of molecular species with th eir surroundings. We show that learned mixing operators mirroring establis hed chemical intuition can reliably predict nonlinear blending behavior\, and introduce a new tool inspired by operator symmetrization to analyze th is interaction.\n\nLastly\, we will demonstrate how advances in ML algorit hm design can enable the accurate prediction of properties which are not p ossible to simulate through conventional means. By adapting techniques fro m computational topology and geometric deep learning\, we construct a nove l neural network architecture capable of predicting protein solubility fro m predicted protein structures. We show that the model’s internal repres entations of protein structure align with those obtained through conventio nal molecular dynamics simulations.\n\nTaken together\, these results refl ect a common organizing principle: the most transferable and physically me aningful models are those whose structures recapitulate the interactions t hat govern chemical behavior.\n\n \; 4:00 pm END:VEVENT END:VCALENDAR