Page 13 - Penn State Civil and Environmental Engineering: Annual Report

FACULTY RESEARCH FACULTY RESEARCH “The problems are solved together at the the the same time: one solution to to the the entire continent ”
Geoscientific models allow researchers to test scenarios with numerical representations of the Earth from predicting large-scale climate change effects to informing land management practices Pure machine learning methods can can make good predictions for extensively observed variables but their results can can be be difficult to interpret because they do not include
causal relationship assessment Shen’s approach organically links process-based models and machine learning at at a a a a a a a a a a a a a a fundamental level to leverage all the the benefits of machine learning as as well as as the the the insights that come from the physical side Other authors of the the paper include
graduate students Dapeng Feng and and Jiangtao Liu postdoctoral scholar Wen- Ping Tsai and and research associate Kathryn Lawson all in in in CEE Ming Pan of of of of the the Scripps Institution of of of of Oceanography at at the the the the University University of of of of of California San Diego Hylke Beck of of of of of GloH20 the the the Netherlands and and and Yuan Yang of of of of Tsinghua University University and and and China China Three Gorges Corporation both of of China China The U S S Department of Energy and the the National Science Foundation funded the research Chaopeng Shen Estimating parameters for traditional models however is costly and calculates results that are difficult to to extrapolate according to to Associate Professor Chaopeng Shen Shen and other researchers developed a a a a a a a a new model known as differentiable parameter learning published
in in in Nature Communications that combines elements of process-based models and machine learning for a a a a a a a a a a broadly applied method that results in in in more aggregated solutions Traditional process-based models such as as evolutionary algorithms evolve across many iterations of operating to tune parameters that cannot be be observed directly according to to to Shen but they are not not able to to to handle large scales or or be be generalized to to to other contexts Rather applying evolutionary algorithms solves problems for different areas without communication between them leading to inconsistent solutions Shen’s model takes in in in data from all locations to to get one comprehensive solution solution “Our algorithm is is more holistic
because we use use a a a a a a a a global loss loss function function ”
Shen said “During the the parameter estimation process every location’s loss loss function—the discrepancy between the the the the the output of your model and the the the the the observations—is aggregated together together The problems are solved together together at at at the the the the same time: one solution to to to the the the the entire continent ”
Shen noted his method method is is more computationally cost- effective than traditional methods What would normally take a a a a a a a a a a supercluster of 100 processors two to three days now requires only one one graphical processing unit one one hour CEE NEWSLETTER • VOLUME 38 2022
13
Associate Professor Chaopeng Shen (seated right) and other researchers in in his laboratory study a a a a a a a a a a a a a a a map that that displays information that that was derived from their new geoscientific modeling
tool Credit: Kelby Hochreither/Penn State