Anuj Karpatne
Virginia Polytechnic Institute and State University, VA·Computer Science·Early-career
Scientists have long predicted how natural systems behave using equations built from theory. These models struggle with complex real-world systems and cannot easily learn from data that theory does not describe well.
- Fields
Computer Science
Earth & Environmental Science
Engineering
- Topics
knowledge-guided machine learning
physics-guided machine learning
scientific machine learning
- Grants
- 3 known grants, through Jun 2028
What they work on
- Develop science-guided curriculum learning and resampling strategies that train deep learning models using both data and scientific equations, so forward predictions, for example in fluid dynamics, stay scientifically consistent.
- Create knowledge-guided inverse models in which knowledge of a scientific forward model guides machine learning to infer the parameters of science-based models.
- Build residual-correcting neural networks that fix systematic biases in the outputs of science-based models while jointly inferring those models' parameters.
- Combine process-based lake models with machine learning (Eco-KGML) to predict water clarity, phytoplankton biomass and hypolimnetic anoxia, oxygen loss in deep water, across thousands of U.S. lakes and reservoirs.
- Use modular compositional learning to identify which water quality processes dominate at each lake, how climate, land use and ecosystem memory interact, and which processes generalize across lake types.
- Develop a physics-guided machine learning framework that uses microscopy images and the physics of cell-exerted forces to predict how cells move, change shape and interact while remodeling fibrous surroundings.
- Extract explainable rules of cell behavior from the physics-guided framework to discover how cells respond to changing physical environments in mechanobiology.
More about the research
Machine learning, which finds patterns directly in data, has a different weakness: when it ignores scientific theory, it has had limited success on scientific problems. This research develops knowledge-guided machine learning, which builds scientific knowledge into machine learning. That knowledge includes equations, simulation models and physical rules. The lab applies this approach to fluid flow, airborne particles (aerosols), lake water quality and the movement of living cells.
Why it matters
Models that combine scientific knowledge with data could give more reliable predictions in fluid dynamics and aerosol science, and in other fields that rely on scientific models. Better water clarity, algae and oxygen predictions for thousands of U.S. lakes could inform federal agencies and the National Ecological Observatory Network (NEON) as climate and land use change. Predicting how cells move could aid studies of embryo development, wound closure, immune response and cancer metastasis.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
CAREER: Unifying Scientific Knowledge with Machine Learning for Forward, Inverse, and Hybrid Modeling of Scientific Systems
NSF · Jul 2023 – Jun 2028
Collaborative Research: MRA: Advancing process understanding of lake water quality to macrosystem scales with knowledge-guided machine learning
NSF · Nov 2022 – Oct 2027
III:Medium:Physics-guided Machine Learning for Predicting Cell Trajectories, Shapes, and Interactions in Complex Dynamic Environments
NSF · Oct 2021 – Sep 2027
Closest labs
Cayelan Carey
Virginia Polytechnic Institute and State University, VA·Ecology & Evolution
Test how far ahead physical, chemical and biological variables in two drinking-water reservoirs can be predicted, by iteratively evaluating real-time forecasts made by competing models against new monitoring data.
Tamer Zaki
Johns Hopkins University, MD·Engineering
Develop evolutional deep neural network (EDNN) algorithms that solve high-dimensional partial differential equations by evolving network parameters according to the governing physics, without relying on training data.
Burak Aksoylu
Texas A&M University-San Antonio, TX·Mathematics & Statistics
Use machine learning to improve iterative local-to-nonlocal coupling, so that peridynamics models only the region where fracture occurs and cheaper local models handle the rest.
The school
Virginia Polytechnic Institute and State University: 35 labs on PI Pages, $657M a year on research, 496 research doctorates a year.
More about Virginia Polytechnic Institute and State University
Reach out
A start for your email. Fill in the brackets.