Physics Guided Machine Learning: A New paradigm for scientific knowledge discovery

09 Oct 2020 - Arun Balajiee

Talk: Physics Guided Machine Learning: A New paradigm for scientific knowledge discovery

Speaker: Xiaowei Jia

Date: 10/09/2020

Today Dr. Jia talked about the idea of using machine learning in modeling physical processes. This is espeically useful in the processes and patterns in nature that only have approximate and possibly inaccurate physics based models. The biggest challenge in modelling natural processes using machine learning is the requirement for a large dataset, inconsistency of the known physical laws for those processes and inability to generalize this in unseen scenarios. The process of his research begins by combining the physics know about physical processes and the information/dataset requirements to build machine learning models. Hence, the goal is to improve the realm of understanding physical processes such as the temperature changes in a lake, if possible, better than the known physics based models for those phenomena. Issues like thermal changes in a lakes and rivers is useful for environmental studies to pattern the changes in climate and ecology of the nearby areas. Hence, if no good physics models exist, studying the ecology could become a challenge. This challenge being solved by Dr. Jia’s work. By introducing a physics based loss function and a dataset built from the data collected by environmental studies of lakes and river ecosystems, Dr. Jia could build an LSTM based model that outperforms an ordinary RNN and also known physics models in predicting the weather changes in a hyperlocal scenario. In the cases of river segment flows, the tributaries that form the rivers could also contribute to the temperature fluxes and hence, Dr. Jia work towards building a graph based RNN which could pre-train on a collective dataset.

Some key takeaways from the talk for me would be:

Building efficient models doesn’t always and necessarily depend on the dataset. sometimes semi-supervised and un-supervised models as Dr. Jia’s could leverage whatever meagre amounts of data is available to process and build a model for prediction
It is really novel idea to be able to aid the research and understanding the physics behind physical processes using Machine Learning. A lot of climate hidden patterns can be discovered in this manner. In the long run, this could save humanity from possible disasters such climate change, hurricanes and other climate related processes. Similarly, understanding the patterns of fluxes in rivers could help in managing better irrigation and agriculture. There are many such use cases where this is applicable
The feature extraction and selection is subjective to the chemical and physical processes of the river and it is not sometimes or entirely possible to find out what could the causal factors for certain phenomena to make predictions
In some cases, these models may have to be pre-trained on simulation datasets to be able to make better predictions. As Dr. Jia’s work shows, this can still build better predictive models that can perform better than the physics based models.