掃碼二維碼觀(guān)看直播

EEH Bilibili直播

EEH 蔻享直播

直播時(shí)間：2023年8月17日 上午9點(diǎn)（北京時(shí)間）

Zoom會(huì)議ID：816 9975 7155

Bilibili鏈接：

https://live.bilibili.com/25002335?broadcast_type=0&;is_room_feed=1&spm_id_from=333.999.0.0（生態(tài)環(huán)境健康EEH）

How to build mechanistically sound machine learing models?

本期主持：

谷成 教授

南京大學(xué)

EEH期刊執(zhí)行主編

? ? ? ? ? ? ??

特邀主講：

Huichun Zhang 教授?

美國(guó)凱斯西儲(chǔ)大學(xué)

? Dr. Huichun (Judy) Zhang is the Frank H. Neff professor in the Department of Civil and Environmental Engineering at Case Western Reserve University. She earned her Ph.D. from Georgia Institute of Technology and her B.S. and M.S. from Nanjing University.?Her research focuses on the fate and transformation of environmental contaminants in natural and engineered aquatic environments and the removal of organic contaminants from contaminated water. Her recent research areas also include predictive modeling for contaminant reactivity and sorption using both classical models and machine learning tools. Dr. Zhang has published in numerous journals, such as Chemical Reviews, Environmental Science and Technology, Water Research, and Applied Catalysis B. She has received seven competitive research grants from the U.S. National Science Foundation as the PI. In addition, Dr. Zhang directed research projects for many other federal and state agencies and industry. She is an Associate Editor for ACS ES&T Water.

報(bào)告摘要

Machine learning (ML) has revolutionized the field of environmental systems modeling by delivering enhanced performance and leveraging diverse input features. However, to ensure the development of robust and meaningful ML models, it is crucial to integrate expert knowledge into the process, particularly in feature selection and post-model interpretation. This presentation aims to illustrate these principles through two compelling examples. In the first case study, we conducted an extensive literature review and compiled a large dataset encompassing the chlorophyll-a index as the output variable. By employing a novel combination of riverine and meteorological features as inputs, we constructed machine learning-based classification and regression models to predict bloom occurrences in Lake Erie. In the second example, we concentrated on developing predictive models for the abiotic reduction of various organic compounds with diverse reducible functional groups, along with the ten most common inorganic compounds, using different Fe(II)-based reductants. To validate these models, we compared the predictions across different chemical groups, reductant identities, and reaction conditions with the known reduction mechanisms. This rigorous evaluation process allowed us to demonstrate the efficacy of our models and their alignment with established scientific principles. In summary, our approach combines expert knowledge, meticulous feature selection, and thorough model interpretation to advance ML modeling in the environmental field.