Chemical, mechanical, electronic, and thermal properties of materials all change at the nanoscale. Entropic fluctuations become more important and materials confined in 1- or 2- dimensions (tubes, wires, sheets) behave differently. Capturing these effects in real devices and applications requires a range of modeling approaches, from hard theory (DFT and kinetics), to soft theory (continuum, statistical mechanics and molecular dynamics), and up through systems engineering approaches.
Applications of include biomedical sensors (nanotube-based optical sensors) and energy (CO2 to fuels, fuel cells, thermal catalysis).
We develop and use machine learning models to solve problems in catalysis. However, it is prohibitively expensive to use full-accuracy DFT to identify catalyst surface with optimal properties due to the sheer number of possible materials, facets, and active sites. The search process can be greatly accelerated by using flexible surrogate models that learn from previous DFT calculations to predict interesting materials. Examples of current efforts include developing new graph convolution structure/property relationships, calibrating uncertainty in these models, and applying these models to new catalyst properties.
We work with a number of experimental collaborators to design and discover catalysts for a range of chemical applications. Examples in thermal catalysis include intermetallics for selective acetylene hydrogenation, metal alloys for propylene epoxidation, and materials for selective direct hydrogenation chemistries. We are also working to develop electrocatalysts to take renewable electricity for the hydrogen/oxygen evolutions reactions, CO2 reduction reaction, and water quality remediation applications. We are also investigating polymer/metal interfaces for fuel cell optimization.
Developing new structure-property relationships or machine learning regression methods requires large self-consistent datasets that are rare in literature. We use and develop modern workflow tools like fireworks and luigi to organize, distribute, and analyze many thousands of DFT calculations. These methods drive down the cost of investigating new chemistries and materials while simultaneously making the results accessible and searchable. This approach also allow for on-line material optimization or active learning approaches to be applied and tested quickly. We are collaborating with Joel Varley (LLNL) to implement workflows for solvation correction in catalysis, Anubhav Jain (LBL) on workflows for water quality applications, and collaborators at University of Toronto and the NRC for mixed metal oxide workflows.
Catalyst design is extremely challenging due to the ovewhelming number of possible materials, surfaces, adsorbates, and configurations. We have built and released large datasets of small molecule intermediates (CO, H, etc) across intermetallic materials to enable and accelerate discovery efforts. Working with collaborators at Facebook AI Research, we built and released the Open Catalyst 2020 (OC20) dataset comprising more than a million relaxations and covering a much wider chemistry space than previously possible. This new dataset will greatly expand the generalizability of ML models.
We build deep learning models to predict catalyst properties from their structure. The diversity of structures and elements make this very difficult. New datasets like the OC20 are also presenting challenges in model training, and we regularly use GPUs to accelerate the training of these models. We showed that graph convolution methods could be used for surface science properties, and are now working with collabrators to improve accuracy and incorporate more physics. We are working to scale methods to thousands of GPUs with the NERSC Early Science Application Program (NESAP) for the upcoming Perlmutter supercomputer.
Choosing the next set of calculations to do or materials to try is extremely challenging. We develop and apply active learning methods to systematically identify the best calculations to do. Efforts include the development of the amptorch code for small-data simulations and custom codes for the active learning process on catalyst surfaces.
We are working with collaborators to apply molecular simulations and AI/ML tools to air quality and wildfire challenges. We are also working with experts at CMU on AI/ML accelerated design of high entropy materials for additive manufacturing.