(2ho) Computational Active Learning of Switchable Materials and Molecular Probes

The central theme of my research interests is in the application and advancement of inverse methods to develop engineered solutions to problems in biophysics, chemistry, and materials science. To this end, I envisage in developing a cross-disciplinary computational design & engineering group working in close collaboration with experimentalists and theorists to address problems with great societal impact. My current interests lie in problems requiring innovative and optimized solutions for enhancing health and reducing illness, and sustainable materials development. Specifically, my projects are designed to tackle the challenges in engineering biosensors, materials for purifying biologics, therapeutics utilizing cellular trash collector, biopolymer properties and biomaterials. This utilizes my combined expertise in molecular dynamics (MD) simulations, enhanced sampling methods, ML and programming developed during my doctoral and postdoctoral training with Prof. Sapna Sarupria at Clemson University and Prof. Andrew L. Ferguson at University of Chicago, respectively.

Teaching Interests

My interest in teaching stems from my own passion for learning and through the inspiration for teaching and research conveyed by all my teachers and mentors. Furthermore, I enjoy teaching because it presents an opportunity to inspire and interact with a new generation of engineers and scientists. I am interested in teaching both undergraduate and graduate level transport phenomena, applied numerical methods, chemical engineering thermodynamics, statistical mechanics and molecular simulations classes. I am well prepared to teach molecular simulations related courses with a focus on its practical applications in materials science, chemical engineering and biophysics. These classes will include topics such as — atomic-scale modeling and simulations, coarse-grained modeling methods, developing force fields, integrator algorithms, thermostat and barostat methods, advanced sampling methods, Markov state models, and recent applications of machine learning in molecular simulations. Furthermore, topics related to scientific computing and best practices in developing robust, scalable and efficient scientific software will be covered. These courses will prepare students for a variety of computational science related careers.

Abstract

Engineering of molecular properties such as in drug discovery, self-assembling materials design, and catalysts development often rely on identifying optimal combination of Hamiltonian parameters or sequence of groups that may be atoms or molecules. To this end, computational active learning comprising molecular dynamics (MD) simulations, enhanced sampling methods and machine learning (ML) has a unique power to efficiently address such problems while gaining insights into the governing forces and mechanisms. We illustrate this using two design studies focused on a) switchable self-assembling materials and b) molecular probes.

Switchable self-assembling materials composed of polarizable nanoparticles (NPs) have several applications such as in energy storage, electronic components and drug delivery. To develop such materials, understanding the design space of NPs is necessary but challenging by conventional computational methods because of the high computational cost. With active learning and by modeling polarization forces using image method, we show the computational cost can be minimized in understanding the design rules for engineering self-assembly of polarizable NPs. The active learning framework includes coarse-grained MD simulations, umbrella sampling method, and ML. We apply it to efficiently map the region of spontaneous and triggerable self-assembly between two polarizable NPs along the experimental design space. The computed map quantitatively predicts the NP assembly and are in good agreement with experimental scattering measurements. Subsequently, we use it to engineer self-assembly of polarizable NPs that are capable of triggered assembly or disassembly via temperature and solvent quality with potential applications in sensors, smart windows, and drug delivery.

As a second example, we will present the application of active learning in efficiently designing molecular probes for developing treatment systems capable of selective removal of recalcitrant water pollutants. We use perfluorooctanesulfonic acid (PFOS) as target recalcitrant water pollutant because of its detrimental effects such as cancer, birth defects and high cholesterol levels in humans. Given the astronomical design space available for molecular probes, conventional computational approaches are hardly tractable. We show the computational cost can be minimized by using active learning involving all-atom MD simulations, parallel-bias metadynamics, deep representational learning of probes and multi-objective Bayesian optimization. We initiate our search by testing the effectiveness of linear probes that can bind via fluorophilic and electrostatic interactions with PFOS. We observed sensitivity of the probes moderately increased with number of fluorinated carbons but their selectivity to PFOS remained low (<-0.25 k_BT) relative to sodium dodecyl sulfate (SDS), a template interferent. Our results show a moderate increase (~0.75 k_BT) in selectivity for the shortest studied probe when one carbon group in a hydrogenated probe is replaced with a primary amine head group. In our presentation, we will discuss the development of the active learning framework using the collected initial data to discover probes with optimal sensitivity and selectivity. The discovered optimal probes will be tested using wet-laboratory experiments and then be deployed for developing efficient water treatment systems.