Artificial Intelligence Projects and Workshops Funded by the Joint Task Force Initiative
Artificially Intelligent Electrochemistry
PIs and Institutions: Chibueze Amanchukwu, UChicago; Rajeev Assary, Argonne
Description: Carbon dioxide emissions are a leading contributor to global climate change, but with the right electrocatalyst and electrolyte, they could instead be a new resource for energy sustainability. The team of Chibueze Amanchukwu, an assistant professor at the Pritzker School of Molecular Engineering, and Rajeev Assary, a staff scientist at Materials Science Division of Argonne, will use AI to find new, efficient ways of “recycling” CO2 into valuable products such as ethanol and ethylene, using clean electricity generated from the sun and wind. The researchers will use quantum chemistry-informed machine learning to sift through computational and experimental data and find new electrolytes that can improve the conversion process. Experiments will then be performed using these obtained chemical insights.
Learned Emulators of Physics Simulations
PIs and Institutions: Rebecca Willett, UChicago; Dana Mendelson, UChicago; Prasanna Balaprakash, Argonne; Jiali Wang, Argonne; Rao Kotamarthi, Argonne
Description: Researchers studying climate, astrophysics, and high-energy physics use large, complex simulations to guide experiments and test theories. These compute-intensive programs can run much faster using emulators that approximate some aspects of those models. Recently, scientists have started creating “learned emulators” using AI neural network approaches, but have not yet fully explored the advantages and potential pitfalls of these surrogates. These aspects will be examined by the team of Rebecca Willett, professor of computer science and statistics at UChicago, Dana Mendelson, assistant professor of mathematics at UChicago, Prasanna Balaprakash, computer scientist at Argonne, Jiali Wang, assistant atmospheric scientist in Environmental Science Division at Argonne, and Rao Kotamarthi, department head of atmospheric science and climate and chief scientist for the Environmental Science Division at Argonne. This team will explore the mathematical limits of these methods and quantify key trade-offs related to accuracy and speed.
Real-Time Adaptive Deep Learning with System-on-Chip Devices for Discovery Science
PIs and Institutions: David Miller, UChicago; Nhan Tran, Fermilab; Andrew A. Chien, UChicago
Description: Machine learning on data is typically performed after it is gathered. But advances in real-time machine learning can analyze data on the fly, allowing scientists to quickly adjust experiments to capture phenomena of interest. That’s particularly appealing to researchers using the Large Hadron Collider at CERN, where a single on-chip system can absorb multiple terabytes of data each second. David Miller, associate professor of physics at UChicago, Nhan Tran, Wilson Fellow at Fermilab, and Andrew A. Chien, professor of computer science at UChicago, will collaborate on the design of new hardware that enables these advanced real-time processes. The resulting “system-on-a-chip” hardware design would help both high-energy physicists and researchers in other data-intensive fields monitor data quality and detect promising results without interrupting the flow of data.
Neural Network Algorithms to Decode the Octopus Neural Network
PIs and Institutions: Peter Littlewood, UChicago; Nicola Ferrier, Argonne; Bobby Kasthuri, UChicago/Argonne
Description: A map of all of the connections in the brain of an advanced animal – together with the means to interpret it – would reveal fundamental principles of organization that would likely revolutionize neuroscience. Using X-ray and electron microscopy imaging at Argonne and UChicago, researchers can now collect the petabyte scale datasets that in principle make this possible.
AI tools will be critical for the analysis. This project will use neural networks to convert three-dimensional raw images into a map — the “connectome” — and make sense of the revealed structure, by building statistical inference of an underlying physical model. These approaches will be applied to the nervous system of the octopus to understand the complex neural network that exists in octopus arms, including chemotactic sensing, camouflage, and motor response. The collaboration brings together animal expertise and husbandry from UChicago and the Marine Biological Laboratory, large-scale instrumentation and imaging from the Advanced Photon Source at Argonne, and high-performance computing and mathematics from Argonne.
Is Climate Change Changing Clouds?
PIs and Institutions: Rebecca Willett, UChicago; Ian Foster, UChicago/Argonne; Elisabeth Moyer, UChicago; Michael Maire, UChicago
Description: Clouds play a dominant role in the Earth’s radiation budget, both reflecting sunlight and trapping infrared radiation. Their responses are the primary source of uncertainty in the numerical simulations used to construct projections of future climate. Satellite data provides detailed imagery of clouds — especially their spatial and temporal distribution. A major open question in climate science is whether researchers can automatically extract spatiotemporal features of clouds from satellite imagery to aid in climate analysis.
This project is a joint effort among Argonne, UChicago Atmospheric Sciences, and UChicago Computer Science aimed at leveraging state-of-the-art computer vision and machine learning tools to extract cloud features from satellite data and characterize the temporal evolution of those patterns in response to a changing climate. This collaborative effort is necessary because off-the-shelf vision tools do not reflect the physical context of cloud formation, necessitating active collaborations between computer science and atmospheric science. Furthermore, the vast quantity of cloud data demands efficient, large-scale, and parallel computing capabilities being developed at Argonne.
Automated Experimental Design for Cosmic Discovery
PIs and Institutions: Brian Nord, Fermilab; Yuxin Chen, UChicago
Description: This project aims to lay the foundations for next-generation cosmic experimental design backed by new artificial intelligence and machine learning methods tailored to cosmology. The collaboration brings together computer scientists and astrophysicists working to understand how data and computation can inform and accelerate cosmic survey design, taking into account previous data, the capabilities of contemporary computational approaches, and insights from the underlying physics.
A distinguishing aspect of the project is its end-to-end focus, going from simulation to real cosmic survey in a closed loop. The work weaves together advances in deep learning, high- dimensional/nonparametric statistics, discrete optimization, and information theory; the developed methodologies will be validated in simulated environments, eventually leading to a new future cosmic experiment which will be evaluated against existing designs by Fermilab.
AI + Measurement
(Eric Jonas & Yuxin Chen, UChicago; Jayakar Thangaraj, Fermilab)
AI can guide scientific experiments before they are started, helping researchers find the best location to aim a telescope or choose which chemical compounds hold the greatest potential as drugs or new materials. Additionally, all modern scientific measurement requires inverting a model of the measurement process to interpret the data, and modern AI techniques have shown great promise in accelerating and improving this reconstruction. The “AI + Measurements” workshop, planned by Eric Jonas and Yuxin Chen, assistant professors of computer science at UChicago, and Jayakar Thangaraj of the lllinois Accelerator Research Center (IARC) at Fermilab, will dig into these opportunities for AI in meteorology, cosmology, particle physics, chemistry and other fields. The workshop will take place at Fermilab and include students and researchers from UChicago, Argonne, and the Toyota Technological Institute at Chicago, brainstorming new collaborations combining cutting-edge AI approaches and experimental science.
AI+Science = CS4All High School Primer Workshop
(Julia Lane, Nick Feamster, and Kyle Chard, UChicago; Michael Papka, Meridith Bruozas and John Domyancich, Argonne; Brian Nord, Fermilab)
For most Chicago Public Schools students, exposure to computer science is limited, and they have little time for advanced topics such as AI or data science for societal issues. This lack of opportunity is exacerbated when students seek internships and other employment experiences and do not have the confidence in their own knowledge to see CS, data science or AI as a possible career pathway for themselves. To address this need, a team of researchers and educators from Argonne, UChicago’s Center for Data and Computing (CDAC) and Fermilab will develop a CS/AI bridge workshop that supports students from Chicago’s South Side community. In the weeklong workshop, to be held twice a year, students will develop a deeper understanding of CS, grow a tangible skillset that is grounded in scientific projects, real-world datasets, and professional tools, and connect with internship opportunities at UChicago, Argonne and Fermilab.