Cloud-resolving model for machine learning buoyant cloudy updraught

Atmospheric model data from simulations carried out using the Unified Model nesting suite u-cj161. This repository includes scripts to perform all the required steps. Retrieval of the model output files from the archive: ml_cape_pdf_1_retrieve_lam_files.py Coarse-graining of the 1.5 km atmospheric model data to a series of different "global model" resolutions by calculating mean profiles of key thermodynamic variables and the calculation of the fraction of that coarse-volume in which the 1.5 km pixels meet the criterion of being bouyant cloudy updraughts: ml_cape_pdf_2d_coarse_grain_bcu.py Subsampling to rebalance the data set since so much of the raw data consists of examples of no convectie activity: ml_cape_pdf_3e_sub_sample_bcu.py There was too much data to process all the raw data in one go, so it was done is several steps. There is then a need to join multiple files together: ml_cape_pdf_3g_stitch_together_bcu_files.py The data is on the Unified Model L70 grid (counting from the bottom up), but the goal is to deploy the machine-learnt algorithm on the E3SM grid (72 levels counting from the top down), so some regirddign is required: ml_cape_pdf_3h_regrid_um_to_e3sm.py Training of 1 two-headed 1d CNN is done using: ml_cape_pdf_4c_bcu_cnn.py Evaluation against witheld data is done using: ml_cape_pdf_5b_validate_BCU.py本数据集源自采用统一模型嵌套套件（Unified Model nesting suite）u-cj161开展的模拟实验所生成的大气模式数据。 本代码仓库包含完成全部必要处理流程所需的脚本文件。 从归档存储中检索模式输出文件的脚本为：ml_cape_pdf_1_retrieve_lam_files.py 将1.5 km分辨率的大气模式数据粗粒度化为多组不同的“全球模式”分辨率：通过计算关键热力学变量的平均廓线，以及统计该粗格体积内满足浮力云上升气流判据的1.5 km像素占比，对应脚本为ml_cape_pdf_2d_coarse_grain_bcu.py 由于原始数据中绝大多数样本均未表现出对流活动，因此需要进行下采样以实现数据集的均衡分布，对应脚本为ml_cape_pdf_3e_sub_sample_bcu.py 原始数据体量过大，无法一次性完成全部处理流程，故采用分步处理方式，后续需将多个拆分文件进行拼接合并，对应脚本为ml_cape_pdf_3g_stitch_together_bcu_files.py 原始数据采用统一模型L70垂直网格（从下向上计数），但机器学习算法最终需部署于E3SM网格（E3SM grid，共72层，从上向下计数），因此需要进行重网格化操作，对应脚本为ml_cape_pdf_3h_regrid_um_to_e3sm.py 训练双头一维卷积神经网络（two-headed 1D CNN）的脚本为ml_cape_pdf_4c_bcu_cnn.py 针对预留验证数据集的模型评估工作采用脚本ml_cape_pdf_5b_validate_BCU.py完成。