11:45 AM - 12:00 PM
[AAS05-11] Development of a coupler h3-Open-UTIL/MP and application
Keywords:Coupling Software, Atmospheric Model, Machine Learning
Since the beginning of climate modeling, coupling software (coupler) has been widely used in climate simulations to exchange information between individual models such as atmosphere and ocean models. Nowadays, couplers are used not only in the meteorology/climate field, but also in various other fields where complex phenomena are simulated. Based on this background, we are developing a general-purpose coupler, h3-Open-UTIL/MP, as a part of the h3-Open-BDEC project.
H3-Open-UTIL/MP can couple any model that meets the following two conditions: 1) it has uniquely numbered and time-invariant grid points, and 2) the time interval of data exchange does not change in time. In addition, it has the following features: 1) ensemble simulation of the coupled models is possible, and 2) it has an API for Python that enables coupling of simulation models written in Fortran with Python applications. Python is a programming language that is widely used in various fields, and it is also used in the field of simulation because of the many libraries developed and provided for data analysis and visualization. In recent years, many machine learning libraries have been developed, and Python has become the main platform used in this field.
In this paper, we introduce one of the applications of h3-Open-UTIL/MP, which is the coupling of the atmospheric model NICAM with the machine learning library PyTorch. In this case study, we trained ML to reproduce the output variables using the input variables for the cloud physics processes of NICAM. As a result, the rough reproduction of the field was good even for learning with a simple three-layer MLP(Multi-Layer Perceptron). However, the extreme values of severe disturbances, which are important for clouds, are under-estimation, and the learning results are not sufficient in terms of reproducing cloud physics. In order to improve the reproductivity, further improvements are needed, such as the use of machine learning algorithms that can handle structured fields and the selection of appropriate variables.
H3-Open-UTIL/MP can couple any model that meets the following two conditions: 1) it has uniquely numbered and time-invariant grid points, and 2) the time interval of data exchange does not change in time. In addition, it has the following features: 1) ensemble simulation of the coupled models is possible, and 2) it has an API for Python that enables coupling of simulation models written in Fortran with Python applications. Python is a programming language that is widely used in various fields, and it is also used in the field of simulation because of the many libraries developed and provided for data analysis and visualization. In recent years, many machine learning libraries have been developed, and Python has become the main platform used in this field.
In this paper, we introduce one of the applications of h3-Open-UTIL/MP, which is the coupling of the atmospheric model NICAM with the machine learning library PyTorch. In this case study, we trained ML to reproduce the output variables using the input variables for the cloud physics processes of NICAM. As a result, the rough reproduction of the field was good even for learning with a simple three-layer MLP(Multi-Layer Perceptron). However, the extreme values of severe disturbances, which are important for clouds, are under-estimation, and the learning results are not sufficient in terms of reproducing cloud physics. In order to improve the reproductivity, further improvements are needed, such as the use of machine learning algorithms that can handle structured fields and the selection of appropriate variables.