Molecular Modeling with Atomistic Machine Learning MethodsAchar, Siddarth Krishnaraja (2024) Molecular Modeling with Atomistic Machine Learning Methods. Doctoral Dissertation, University of Pittsburgh. (Unpublished) This is the latest version of this item.
AbstractQuantum mechanics (QM) provides a method for computing energies and forces on atoms and ions within condensed phase and molecular systems to arbitrary accuracy. This allows us, in principle, to compute a vast array of properties, including reaction pathways, equilibrium properties, and transport properties. However, application of QM is severely limited by the adverse scaling of QM methods. In practice, QM-based simulations are limited to a few hundred atoms with dynamic time scales of 10s to 100s of ps. Machine learning (ML) potentials can be used to effectively extend the range of QM methods by allowing us to compute energies and forces of condensed phase and molecular systems with near-QM accuracy for tens of thousands of atoms with dynamic time scales to tens of ns. We focus on developing ML potentials for various applications. Firstly, we trained highly-accurate deep learning potentials (DPs) for graphanol (hydroxy functionalized graphane). Our simulations demonstrated that graphanol conducts protons efficiently without hydration. Our investigations into proton diffusion and barriers, along with temperature fluctuations, revealed insights for designing improved proton exchange membranes. Additionally, we employed accurate DPs for modeling diffusion in metal-organic frameworks (MOFs) like UiO-66, and interface diffusion in chalcogenide alloys and electrodes for non-volatile memory cells, using the moment tensor approach (MTP) for construction. ML potentials do not contain information about charge densities. Therefore, we developed a method called DeepCDP (deep learning for charge density prediction) to predict electron densities solely from atomic coordinates. We achieved linear scaling in density prediction with near-DFT accuracy. Our predicted densities were utilized to track protons in graphanol and compute dipole moments for several molecules. Share
Details
Available Versions of this Item
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|
![[feed]](/images/feed-icon-32x32.png){kind=icon}