Past Seminars

Quin Hu "Machine learning for bond breaking: Training CASPT2 level neural network potentials for carbon carbon dissociation in alkanes"

Quin Hu
Graduate Student, Goodpaster Group

Neural Network potentials are developed which accurately make and break bonds for use in molecular simulations. Large, condensed phased, and extended systems remain a challenge for theoretical studies due to the compromise between accuracy and computational cost in calculations. Machine learning methods provide a new approach to solve this trade off by leveraging large datasets to train on highly accurate calculations on small molecules and then extending learned features to predict larger systems. Specifically, neural network potentials have shown great promise to achieve high accuracy calculations with low computational cost. In this project, we introduce a novel method to combat the accuracy versus cost battle in computational chemistry.

We designed a training algorithm to generate neural network potentials to predict molecular energies of specific systems of interest; this method aims to use the minimum amount of data of smaller/sub systems to train neural network potentials to the accuracy level of advanced ab initio methods. Starting with a neural network potential trained at the density functional theory (DFT) level, we generate homolytic carbon-carbon bond dissociation data of small size alkanes with DFT energies to train the potentials to accurately predict bond dissociation at the DFT level. Then, using transfer learning, we retrained the neural network potential to complete active space second-order perturbation theory (CASPT2) level of accuracy. We demonstrate that the neural network potential only requires bond dissociation data of a few small alkanes to accurately predict bond dissociation energy in larger alkanes. This training algorithm can be further applied to any type of bond or any level of theory.

 

Huakun "Quin” Hu

Josh Kretchmer "Density matrix embedding theory methods for non-equilibrium electron dynamics in extended systems"

Josh Kretchmer
Assistant Professor, Georgia Tech

The simulation of non-equilibrium electron dynamics in extended systems provides a challenge for theoretical methods due to the need to treat both large system sizes and electron correlation. In this talk I will present our work developing extensions of density matrix embedding theory (DMET) methods to treat real-time dynamics. In DMET, the total system is partitioned into an impurity corresponding to a region of interest coupled to the surrounding environment, which is efficiently represented by a quantum bath of the same size as the impurity. We take advantage of this partitioning to develop equations of motion for the coupled impurity and bath embedding problem that allow for efficient and accurate simulation of real-time dynamics. The methodology is able to simulate non-equilibrium electron dynamics in the presence of strong correlation, reaching total system sizes unobtainable by conventional methodology.

Josh Kretchmer

Amarda Shehu "AI-enabled Discovery of Macromolecular Structure, Dynamics, and Function"

Amarda Shehu
Professor, George Mason University

Biology has undergone many disruptions and revolutions that have opened or redirected entire domains of scientific enquiry. Anfinsen showed us that protein tertiary structure was largely encoded in the amino-acid sequence. John Kendrew’s famous sentences “The way in which the chain of amino acid units in a protein molecule is coiled and folded in space has been worked out for the first time. The protein is myoglobin, the molecule of which contains 2,600 atoms.” instigated decades of computational studies on macromolecular structure, dynamics, and function, starting with the seminal work of Scheraga, Karplus, Levitt, Warshel, and others. Some of the most interesting computational concepts and techniques were debuted in these studies, many of which we would now categorize as “artificial intelligence” (AI). In my own work, I focused heavily, and still do, on the question of knowledge representation, which is central to AI. Then came the data revolution and the revival of neural networks. What was old became new. My laboratory showed data-driven models to be more powerful in many respects, yet not always satisfying. And now comes news DeepMind’s Alphafold2 has solved a 50-year grand challenge in biology. Many of us wonder what these increasingly more frequent AI disruptions mean for our disciplines, our students, and our careers. I will conclude this talk with my admittedly biased perspective and argue that, while we still need to figure out a new collaborative medium, these AI disruptions provide wonderful opportunities to get into complex, messy, integrative scientific enquiries.

Speaker Bio
Dr. Amarda Shehu is a Professor in the Department of Computer Science in the Volgenau School of Engineering with affiliated appointments in the Department of Bioengineering and School of Systems Biology at George Mason University. She is also Co-Director of the Center for Advancing Human-Machine Partnerships (CAHMP), a Transdisciplinary Center for Advanced Study at George Mason University. Shehu obtained her Ph.D. in Computer Science from Rice University in 2008, where she was also an NIH pre-doctoral fellow. Shehu's research focuses on novel algorithms in artificial intelligence and machine learning to bridge between computer and information sciences, engineering, and the life sciences. In particular, her laboratory has made many contributions in bioinformatics and computational biology regarding the relationship between macromolecular sequence, structure, dynamics, and function. Shehu has published over 130 technical papers with postdoctoral, graduate, undergraduate, and high-school students. She is the recipient of an NSF CAREER Award, and her research is regularly supported by various NSF programs, as well as state and private research awards. Shehu is also the recipient of the 2018 Mason University Teaching Excellence Award, the 2014 Mason Emerging Researcher/Scholar/Creator Award, and the 2013 Mason OSCAR Undergraduate Mentor Excellence Award. She currently serves as Program Director at the National Science Foundation in the Information and Intelligent Systems Division of the Computer and Information Science and Engineering Directorate.

Amarda Shehu

Peter Wirnsberger "Optimizing Free Energy Estimation with Machine Learning"

Peter Wirnsberger
Senior Research Scientist, DeepMind

Computer-based free energy estimation has been an active field of research for decades, with many successful applications in physics, materials science and biology. Free energy perturbation (FEP) [1] is a bedrock technique for estimation of free energy differences. A requirement for fast and reliable convergence of the estimator, however, is that the respective distributions share a large overlap in configuration space. An elegant strategy to address this problem is to increase overlap using configuration space mappings. This approach, known as Targeted Free Energy Perturbation (TFEP) [2], has the appealing property that it can achieve immediate convergence when combined with a perfect mapping. Defining a good high-dimensional mapping, however, is a challenging task. 

In this talk, I will present our recent work [3] in which we turn TFEP into a machine learning problem. In our approach, we represent the mapping by a deep neural network whose parameters are optimized so as to maximize overlap. We test our targeted estimators on a prototypical solvation system for which we generate training data using molecular dynamics simulations. Our neural network architecture respects two important symmetries of the system, namely periodic boundary conditions and permutational symmetry of identical particles. We observe that our technique leads to an improved accuracy of the free energy estimates compared to baselines, without requiring any additional data for training the network.

References
[1] R. W. Zwanzig, J. Chem. Phys. 22, 1420 (1954).
[2] C. Jarzynski, Phys. Rev. E 65, 046122 (2002).
[3] This research by P. Wirnsberger, A.J. Ballard, G. Papamakarios, S. Abercrombie, S. Racanière, A. Pritzel, D. Jimenez Rezende, and C. Blundell has been first published in J. Chem. Phys. 153 , 144112 (2020), with the permission of AIP Publishing. PW and AJB contributed equally to this work.

Speaker Bio
Peter is a Senior Research Scientist at DeepMind. His research interests include developing new machine learning algorithms as well as applying them to problems that lie at the intersection of machine learning and physics. Peter obtained his PhD in Theoretical Chemistry from the University of Cambridge, where he worked with Daan Frenkel on quantifying polarization effects in molecular systems far from thermal equilibrium.

Peter Wirnsberge

Sam Schoenholz "JAX MD: A Framework for Differentiable Physics"

Sam Schoenholz
Google Brain

I will talk about JAX MD, a software package for performing differentiable physics simulations with a focus on molecular dynamics. JAX MD includes a number of physics simulation environments, as well as interaction potentials and neural networks that can be integrated into these environments without writing any additional code. Since the simulations themselves are differentiable functions, entire trajectories can be differentiated to perform meta-optimization. These features are built on primitive operations, such as spatial partitioning, that allow simulations to scale to hundreds-of-thousands of particles on a single GPU. My talk will include an introduction to automatic differentiation through the JAX software package (www.github.com/google/jax) so no background is required. If you are interested in trying out JAX MD, it is available at github.com/google/jax-md

Speaker Bio
Sam is a Senior Research Scientist at Google Brain working at the intersection between Machine Learning and Physics. His work focuses on better understanding neural networks using techniques from statistical physics as well as applying advances in Machine Learning to physical systems. Sam received his PhD from the University of Pennsylvania where he used machine learning to study disordered materials and glassy liquids.

Sam Schoenholz

Johannes Klicpera "Graph Neural Networks for Chemistry: Building Powerful Models Without Breaking the Laws of Physics"

Johannes Klicpera
TU Munich

Recent advances in machine learning for Chemistry promise to speed up molecular property prediction, drug screening, inverse synthesis planning and other tasks by multiple orders of magnitude. Many of these recent advances are based on graph neural networks, which have established themselves as a powerful model for molecular and physical systems. By directly modeling the interactions between particles they originally promised to replace the sophisticated hand-crafted features prevalent in classical machine learning models (e.g. kernel ridge regression). However, while graph neural networks do indeed outperform classical methods on large benchmarks, they still struggle with generalization on tasks with limited available data.

In this talk I will first give a general overview of graph neural networks and their underlying principles and limitations. I will then focus on their application to molecular dynamics and present modern approaches that try to marry the best of both worlds by building physical knowledge directly into the model [1]. Doing so significantly improves both their overall performance and their generalization capabilities, and thus opens venues for research on molecules far away from the equilibrium [2].

References
[1] Johannes Klicpera, Janek Groß, Stephan Günnemann. Directional Message Passing for Molecular Graphs. ICLR 2020.
[2] Johannes Klicpera, Shankari Giri, Johannes T. Margraf, Stephan Günnemann. Fast and Uncertainty-Aware Directional Message Passing for Non-Equilibrium Molecules. ML4Molecules workshop, NeurIPS 2020.

Speaker Bio
Johannes Klicpera is a PhD student in the Data Analytics and Machine Learning group lead by Stephan Günnemann at TU Munich. His research focuses on machine learning for relational data, from web-scale social networks to small molecules. Before starting his PhD he studied Computer Science and Physics at TU Munich and the University of Cambridge. More information on his research is available in his Google Scholar profile.

Johannes Klicpera

Bingqing Cheng "Computing thermodynamic and transport properties using machine learning potentials"

Bingqing Cheng
University of Cambridge

A central goal of computational physics and chemistry is to predict material properties using first principles methods based on the fundamental laws of quantum mechanics. However, the high computational costs of these methods typically prevent rigorous predictions of macroscopic quantities at finite temperatures, such as chemical potential, heat capacity and thermal conductivity.

In this talk, I will first discuss how to enable such predictions by combining advanced statistical mechanics with data-driven machine learning interatomic potentials. As an example [1], for the omnipresent and technologically essential system of water, a first-principles thermodynamic description not only leads to excellent agreement with experiments, but also reveals the crucial role of nuclear quantum fluctuations in modulating the thermodynamic stabilities of different phases of water. As another example [2], we simulated the high pressure hydrogen system with converged system size and simulation length, and found, contrary to established beliefs, supercritical behaviour of liquid hydrogen above the melting line. Besides the computation of thermodynamic properties, I will talk about transport properties: Ref [3] proposed a method to compute the heat conductivities of liquid just from equilibrium molecular dynamics trajectories.
 
During the second part of the talk, I will rationalize why machine learning potentials work at all, and in particular, the locality argument. I'll show that a machine-learning potential trained on liquid water alone can predict the properties of diverse ice phases, because all the local environments characterising the ice phases are found in liquid water [4].
 
References
[1] Bingqing Cheng, Edgar A Engel, Jörg Behler, Christoph Dellago, Michele Ceriotti. (2019) ab initio thermodynamics of liquid and solid water. Proceedings of the National Academy of Sciences, 116 (4), 1110-1115.
[2] Bingqing Cheng, Guglielmo Mazzola, Chris J. Pickard, Michele Ceriotti. (2020) Evidence for supercritical behaviour of high-pressure liquid hydrogen. Nature, 585, 217–220
[3] Bingqing Cheng, Daan Frenkel. (2020) Computing the Heat Conductivity of Fluids from Density Fluctuations. Physical Review Letters, 125, 130602
[4] Bartomeu Monserrat, Jan Gerit Brandenburg, Edgar A. Engel, Bingqing Cheng. (2020) Liquid water contains the building blocks of diverse ice phases. Nature Communications 11.1: 1-8.


Speaker Bio
I am a Departmental Early Career Fellow at the Computer Science Department, University of Cambridge. My research uses computer simulations to understand and predict material properties, with a particular focus on exploiting machine-learning methods to extend the scope of atomistic simulations. I did my PhD in Michele Ceriotti's group, at EPFL.

Triangular Molecular Magnets: Chemical Models for Protons, Qubits, or Quantum Sensors?

Triangular Molecular Magnets

Mark Pederson
Professor, Department of Physics
The University of Texas at El Paso

The size range and deceptive complexity albeit behavioral simplicity of molecular magnets attracts physical scientists from many disciplines and challenges them to understand how 50-100 nuclei and 200-1000 electrons can exhibit such simple collective behavior. For example, quantum tunneling of magnetization,  which occurs in broken-spin-symmetry magnetic molecules illustrates the power of density-functional-based pictures for predicting both the  magnetic strength of molecules and the magnetic fields at which quantum tunneling occurs.  Alternatively, the spin-electric effect[1-4] explicitly challenges the notion that single-determinantal theories can describe the physics leading this phenomenon. However, the large molecular size resists quantitative quantum chemical explanations and a combination of model Hamiltonians with density-functional treatments are the optimal means for exploring these intrinsically multi-configurational problems. I will review previous work on the Cu3 molecular magnet and show how the combination of broken symmetry density-functional theory, with simple self-interaction corrections and spin-orbit inclusion, can be used to derive three-spin Heisenberg Hamiltonians that describe the Dzyaloshinskii-Moriya induced splitting of degenerate low-energy Kramer doublets into S=1/2 chiral and anti-chiral pairs.  The resulting energy level diagrams will be compared to that of a three-quark system.

Triangular Molecular Magnets

The second half of this talk features the Fe3O(NC5H5)3(O2CC6H5)6 molecule[4] that is the first possible spin-electric system based upon spin 5/2.centers. As a curiosity, I discuss the rather unusual point-group symmetry, which includes a set of rotation matrices that are hauntingly similar to those that appear in elementary particle wavefunctions,. We call the generator of these rotations matrices (above) RQ[2]. Using standard density-functional methods we show that the spin-electric behavior of this molecule could be more interesting due to energetically competitive reference states with high and low local spins (S=5/2 vs. S=1/2) on the Fe3+ ions. We provide spectroscopies to deduce the presence of both states and note that similar multiferroic behavior exists in the Mn3 molecular magnet[3].  Rationale for use of a new version to self-interaction corrections, FLOSIC, to improve quantitative predictions, especially in lanthanide systems and periodic systems will be included.[5]

Supported by the M2QM Energy Frontier Research Center, Grant #: DE-SC0019330 (Molecular Magnetism), Grant # DE-SC0018331 (FLOSIC), Texas STARS, & the C. Sharp Clark Cook Physics Chair Endowment.

[1] M. F. Islam, J. F. Nossa, C. M. Canali & MRP, First-principles study of spin-electric coupling in a Cu3 single molecular magnet, PRB 82 155446 (2010).

[2] A.I. Johnson, M.F. Islam, C.M. Canali & MRP, A Multiferroic molecular magnetic qubit, JCP. 151, 174105 (2019).

[3] Z. Hooshmand & MRP, Control of spin-ordered Mn3 Qubits: A density-functional study, Physical Chemistry and Chemical Physics (2020)

[4] A. K. Boudalis, J. Robert & P. Turek, 1st demonstration of magnetoelectric coupling in a polynuclear molecular nanomagnet via EPR studies  Fe3O(O2CPh)6(Py)3ClO4, Chem. Eur. J 24 14896-14900 (2018).

[5] MRP, A. Ruzsinszky and J.P. Perdew, Communication: Self-Interaction correction with unitary invariance in density functional theory, J.Chem. Phys. 140 121105 (2014)

Nano-diving into the clouds: Uncovering molecular mechanisms of heterogeneous ice nucleation

Sapna Sarupria
Associate Professor
Clemson University

The presence of particles such as dust and pollen affect cloud microphysics significantly through their effect on the state of water. These particles can hinder or accelerate the liquid-to-solid transition of water, and also affect the ice polymorph formed in the clouds. This indirectly cloud reflectivity, cloud lifetime, and precipitation rates. While a predominant phenomenon, the understanding of the surface factors that affect ice nucleation is minimal. In our research, we use molecular simulations to illuminate the pathways through which surface properties influence ice nucleation. Experiments cannot probe the length and time scales relevant to nucleation. While molecular simulations, in principle, can probe the length and time scales of nucleation, in practice nucleation is challenging to sample. Nucleation is often associated with large free energy barriers and thus, is difficult to sample in straightforward simulations. Advanced sampling techniques and other creative approaches are needed. In this talk, I will discuss the insights we have obtained on heterogeneous ice nucleation through studies of three surfaces – silver iodide, kaolinite and mica. I will also highlight the synergistic combination of experiments and simulations in understanding heterogeneous ice nucleation. I will introduce a recently developed method in our group facilitate computational studies of heterogeneous nucleation. I will conclude by providing a perspective on the broader implications of our studies on interfacial phenomena and surface design.

Speaker Bio 
Dr. Sapna Sarupria is an associate professor in Chemical and Biomolecular Engineering at Clemson University. She received her Masters’ from Texas A & M University where her thesis focused on thermodynamic modeling of clathrate hydrates of gas mixtures formed in the presence of electrolyte solutions. She obtained her Ph.D. from Rensselaer Polytechnic Institute, where she studied pressure effects on water-mediated interactions and proteins. She was a postdoctoral researcher in Princeton University and studied hydrate and ice nucleation using advanced path sampling techniques. She received the NSF CAREER award, ACS COMP Outstanding Junior Faculty Award and Clemson’s Board of Trustees Award of Excellence. She is an active member of Women in Chemical Engineering (WIC) and Computational and Molecular Science and Engineering Forum (CoMSEF) in AIChE.

Research Overview
Sarupria group research focuses on surface-driven phenomena. Current projects include heterogeneous ice nucleation, protein adsorption on surfaces and fouling on water purification membranes. The central theme in Sarupria group involves developing cutting-edge sampling techniques in molecular simulations and applying them in understand long standing problems in condensed matter. We recently developed novel transition path sampling methods and software to enable their largescale implementation in HPC infrastructure. These methods will be used to study ice nucleation, and reactions in condensed phases including enzymatic reactions.

 

Sapna Sarupria

QM-ORSA: An accurate computational protocol to explore the tip of the antioxidant iceberg

Annia Galano
Professor
Metropolitan Autonomous University

Chemical antioxidants are potential candidates to ameliorate the deleterious effects of oxidative stress-related diseases. However, both oxidative damage and antioxidant protection are complex and interrelated processes. That is why investigations on this field frequently focus on specific aspects of the whole phenomenon. Probably, the most widely explored aspect of chemical antioxidants is their ability to deactivate free radicals. The Quantum Mechanics-based Test for Overall Free Radical Scavenging Activity (QM-ORSA) is a computational protocol designed to be a reliable tool in the study of radical-molecule reactions in solution. It can be used to provide a universal and quantitative way of evaluating the free radical scavenging activity of chemical compounds. It provides a separated quantification of such activity in polar (aqueous) and non-polar (lipid) media. It includes two different scales for quantification: (i) absolute, based on overall rate coefficients; and (ii) relative, using Trolox as a reference antioxidant. QM-ORSA also allows identifying the main mechanisms of reaction involved in the free radical scavenging activity of chemical antioxidants and establishing the influence of pH on such an activity. The QM-ORSA protocol has been validated versus experimental results, and its uncertainties were proven to be no larger than those arising from experiments.

  • Annia Galano, Juan Raúl Alvarez-Idaboy “A Computational Methodology for Accurate Predictions of Rate Constants in Solution: Application to the Assessment of Primary Antioxidant Activity” J. Comput. Chem. 2013, 34 (28), 2430–2445.
  • Annia Galano, Gloria Mazzone, Ruslán Alvarez-Diduk, Tiziana Marino, J. Raúl Alvarez-Idaboy, Nino Russo. “Food Antioxidants: Chemical Insights at the Molecular Level” Annu. Rev. Food Sci. Technol. 2016, 7, 335–352.
  • Annia Galano, Juan Raúl Alvarez-Idaboy “Computational strategies for predicting free radical scavengers’ protection against oxidative stress: Where are we and what might follow?” Int. J. Quantum Chem. 2019, 119, e25665 (23 páginas).

Speaker Bio
Professor Galano received her bachelor’s degree and her doctorate at Havana University, Cuba. She performed postdoctoral studies at the Mexican Institute of Petroleum and at the Metropolitan Autonomous University (UAM, according to its Spanish acronym). She had research stays at Uppsala University and Calabria University and joined the faculty at the UAM in 2008.

Research Overview
Professor Annia Galano’s research group applies computational chemistry to the investigation of physicochemical insights directly related to atmospheric pollution, carbon-based nanomaterials, oxidative stress, and antioxidant activity. The current central themes of her group are:

(i)    The elucidation of the chemical behavior of antioxidants and the relative efficiency for that purpose of a large variety of chemical compounds.
(ii)    The systematic design of new antioxidants with multifunctional behavior and possible neuro-protective effects.

Her group is involved in interdisciplinary research that mixes theoretical chemistry, organic synthesis, analytical chemistry, chemical biology, and medicinal chemistry.

Information: www.agalano.com
ORCID: 0000-0002-1470-3060
E-mail: agal@xanum.uam.mx, annia.galano@gmail.com 
 

Annia Galano