GSA Connects 2021 in Portland, Oregon

Paper No. 245-12
Presentation Time: 4:20 PM

USING A PHYSICS-BASED MODEL GUIDED NEURAL NETWORK TO PREDICT SOIL REACTION FRONTS


WEN, Tao, Department of Earth and Environmental Sciences, Syracuse University, Syracuse, NY 13244, CHEN, Chacha, College of Information Science and Technology, Pennsylvania State University, University Park, PA 16802, ZHENG, Guanjie, John Hopcroft Center for Computer Science, Shanghai Jiao Tong University, Shanghai, 200240, China, BANDSTRA, Joel, Department of Mathematics, Engineering, and Computer Science, Saint Francis University, Loretto, PA 15940 and BRANTLEY, Susan L., Earth and Environmental Systems Institute, Pennsylvania State University, University Park, PA 16802; Department of Geosciences, Pennsylvania State University, University Park, PA 16802

Numerous physics-based models (PBMs) have been proposed to model the soil reaction fronts to study soil formation and other surface processes. These PBMs are often highly parameterized. In this study, we developed a hybrid neural network (HNN) that integrated the neural network (NN) into the PBM to simulate the geochemical profile of feldspar contents in soils. The developed HNNs used six environmental variables known from the PBMs that were important to characterize the soil reaction fronts, including site climate characteristics (temperature and precipitation), geomorphic parameters (soil residence time and erosion rate), and parent material mineralogy (quartz content and albite content of the feldspar). The NN was introduced to minimize the difference between predicted results from PBMs and field observations. For each of the combinations of environmental variables, we trained and tested an HNN and a PBM for three different sets of soil training and test data. To seek the best performing HNN in terms of mimicking the corresponding PBM prediction results, we evaluated the percent difference in MSE between each of the HNNs and (1) any HNNs using a subset of predictor variables used in the former HNN and (2) the corresponding PBM in the training phase. Among those best-performing HNNs, soil age was most frequently included in the HNNs indicating that it was the most useful predictor variable to construct an HNN that can achieve the prediction accuracy comparable to the PBM. Instead, precipitation is the least useful predictor variable in that regard. In addition, the preliminary results also showed that at least two predictor variables were needed for an HNN to achieve an MSE within 1% different from that of the corresponding PBM. The best performing HNNs were also used to predict the reaction front for soil profiles not included in the corresponding training dataset. Prediction results showed that HNNs can generally mimic the simulated reaction front from the PBM with respect to the slope of the reaction front rather than the depth of the reaction front. This is the first time a NN was incorporated into a PBM to develop an HNN to simulate the depth profile of soil geochemistry.