[{"data":1,"prerenderedAt":202},["ShallowReactive",2],{"content-query-5FsQPovcUR":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":196,"_id":197,"_source":198,"_file":199,"_stem":200,"_extension":201},"/technology-blogs/en/2997","en",false,"","Idea Sharing: Introduction to AI-GOMS, an Ocean-Based Foundation Model","Weather and ocean are deeply interconnected. The heat exchange between ocean circulation and atmosphere has a great impact on weather changes.","2023-11-13","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/03/01/a8baa177c50e471080f0da3d0a80b03d.png","technology-blogs",{"type":14,"children":15,"toc":193},"root",[16,24,34,42,47,52,57,65,70,75,83,88,93,98,103,110,115,123,128,136,149,160,171,182],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"idea-sharing-introduction-to-ai-goms-an-ocean-based-foundation-model",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28],{"type":17,"tag":29,"props":30,"children":31},"strong",{},[32],{"type":23,"value":33},"Author: Yu Fan Source: Zhihu",{"type":17,"tag":25,"props":35,"children":36},{},[37],{"type":17,"tag":29,"props":38,"children":39},{},[40],{"type":23,"value":41},"Background",{"type":17,"tag":25,"props":43,"children":44},{},[45],{"type":23,"value":46},"Weather and ocean are deeply interconnected. The heat exchange between ocean circulation and atmosphere has a great impact on weather changes. Currently, a large number of AI models have demonstrated their accuracy in medium-range weather forecasting, for example, GraphCast[1] and Pangu-Weather[2]. In terms of nowcasting, DeepMind took the first step to use generative models for rainfall forecasting[3], while Long Mingsheng's team from Tsinghua University proposed NowcastNet[4] with physical mechanisms. We believe that the ocean will be the next arena for AI in the meteorological field. Recently, Huang Xiaomeng's team from Tsinghua University worked with Huawei Advanced Computing and Storage Laboratory and proposed AI-GOMS[5], a pioneering ocean-based foundation model.",{"type":17,"tag":25,"props":48,"children":49},{},[50],{"type":23,"value":51},"There are similarities and differences between ocean-based forecasting and atmospheric variable forecasting. Ocean-based weather forecasting employs AI similarly to medium-term forecast models and relies on reanalysis data. ERA5 is commonly used for atmospheric reanalysis. It also provides the data of weather experiments for foundation models such as Pangu. There are also ocean-based models such as HYCOM. Therefore, it is feasible to port part of the medium-term forecast model to ocean-based forecasting.",{"type":17,"tag":25,"props":53,"children":54},{},[55],{"type":23,"value":56},"However, the physical mechanisms between the ocean and atmosphere are different. To integrate physical mechanisms into the model, we need to adjust the model accordingly. In addition, beside physical characteristics including the sea surface temperature and wave height, biochemical-related variables, such as nitrate and chlorophyll concentrations, should also be forecasted. In other words, different strategies in the downstream tasks are required. Currently, AI exploration in oceans is still in the early stage, leaving vast opportunities for foundation models. This blog introduces the recent ocean-based AI-GOMS[5] model.",{"type":17,"tag":25,"props":58,"children":59},{},[60],{"type":17,"tag":29,"props":61,"children":62},{},[63],{"type":23,"value":64},"1. AI-GOMS",{"type":17,"tag":25,"props":66,"children":67},{},[68],{"type":23,"value":69},"AI-GOMS is an autoencoder foundation model based on Fourier operators. Using the basic variables and boundary conditions of the ocean pattern in the initial field as the input, its backbone model forecasts five types of basic ocean variables, including the sea temperature (T), sea salinity (S), zonal velocity (U), meridional velocity (V), and sea surface height (SSH). AI-GOMS uses HYCOM global reanalysis data as the input to perform a day-level global forecast for the next 30 days.",{"type":17,"tag":25,"props":71,"children":72},{},[73],{"type":23,"value":74},"Figure 1 shows the AI-GOMS model architecture. Because in the ocean mode the input is grid data, the author performs an operation similar to Vision Transformer (ViT) to divide 2D grid data into non-overlapping patches and convert them into 1D sequence tokens for subsequent encoding. According to Figure 1 (a), we can see that the model randomly masks the input in order to learn intrinsic features of the data and alleviate overfitting to obtain a better long-term prediction effect. The tokens after random mask are mapped to the frequency domain through the Fourier modules. The Fourier modules, designed based on the adaptive Fourier neural operator (AFNO) practice, features the advantage of variable lengths of input sequences, which enables the previous random mask operation. The decoder module filters the features output by the Fourier modules in the frequency domain to select the variable features helpful to the forecast. Then a basic ocean variable prediction is made after projection.",{"type":17,"tag":25,"props":76,"children":77},{},[78],{"type":17,"tag":79,"props":80,"children":82},"img",{"alt":7,"src":81},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/03/01/a14ba87e1be24c40b4350848095a423d.png",[],{"type":17,"tag":25,"props":84,"children":85},{},[86],{"type":23,"value":87},"Figure 1 AI-GOMS model architecture",{"type":17,"tag":25,"props":89,"children":90},{},[91],{"type":23,"value":92},"Based on the basic ocean variable prediction of the backbone model, three downstream lightweight forecast tasks can be performed in the downstream module: biochemical variable prediction, downscaling, and wave prediction. Figure 1 (b) shows the design of the downstream module. The backbone model starts from the input of Figure 1 (a) and ends with the decoder module. The output of the backbone model is used as the feature tensor.",{"type":17,"tag":25,"props":94,"children":95},{},[96],{"type":23,"value":97},"In the backbone model, the feature tensor is the value of the basic variables output through projection. Whereas in the downstream module, the feature tensor is combined with the downstream input, and the downstream output is obtained through a lightweight fine-tuning network. Downstream input and the corresponding lightweight fine-tuning network varies with different downstream tasks. The five variables of Kuroshio at 1/4° spatial resolution are used as the downscaling input. Through the residual convolution blocks and upscaling ConvTranspose2d layer, it outputs Kuroshio grid data at 1/12° high spatial resolution. For the biochemical variable and wave prediction, 1D-AFNO operators and a projection layer are used as the fine-tuning network.",{"type":17,"tag":25,"props":99,"children":100},{},[101],{"type":23,"value":102},"The initial condition extracted from ERA5 data provides the downstream input required for wave prediction, while biochemical variable prediction uses satellite assimilation data provided by NASA to forecast eight biochemical variables (total chlorophyll a concentration, chlorophyte concentration, diatom concentration, coccolithophores concentration, cyanobacteria concentration, iron concentration, nitrate concentration, and mixed layer depth. Figure 2 shows the forecast effects of the three downstream tasks.",{"type":17,"tag":25,"props":104,"children":105},{},[106],{"type":17,"tag":79,"props":107,"children":109},{"alt":7,"src":108},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/03/01/6249920585c84c49bc2686e5eeb62d50.png",[],{"type":17,"tag":25,"props":111,"children":112},{},[113],{"type":23,"value":114},"Figure 2 Prediction results of AI-GOMS downstream tasks (a: downscaling result in Kuroshio , taking the sea surface stream velocity as an example; b: wave height; c: chlorophyll concentration)",{"type":17,"tag":25,"props":116,"children":117},{},[118],{"type":17,"tag":29,"props":119,"children":120},{},[121],{"type":23,"value":122},"2. Conclusion",{"type":17,"tag":25,"props":124,"children":125},{},[126],{"type":23,"value":127},"Foundation model application in ocean forecasting is still in the initial stage. As improved ViT models and generative models can be applied in this field, AI-GOMS has made a preliminary attempt. In the future, the combination of ocean physical mechanisms and deep learning can become a meaningful research topic. In addition, ocean observations are more dependent on satellite data to avoid the high cost of in situ observation. With China's FY satellites, integration with satellite, ocean forecasts and foundation models will mark a significant future direction, and bring great economic benefits in practice.",{"type":17,"tag":25,"props":129,"children":130},{},[131],{"type":17,"tag":29,"props":132,"children":133},{},[134],{"type":23,"value":135},"References",{"type":17,"tag":25,"props":137,"children":138},{},[139,141],{"type":23,"value":140},"[1]Remi Lam, Alvaro Sanchez-Gonzalez, et al. GraphCast: Learning skillful medium-range global weather forecasting. ",{"type":17,"tag":142,"props":143,"children":147},"a",{"href":144,"rel":145},"https://arxiv.org/abs/2212.12794",[146],"nofollow",[148],{"type":23,"value":144},{"type":17,"tag":25,"props":150,"children":151},{},[152,154],{"type":23,"value":153},"[2]Bi, K., Xie, L., Zhang, H. et al. Accurate medium-range global weather forecasting with 3D neural networks. Nature 619, 533–538 (2023). ",{"type":17,"tag":142,"props":155,"children":158},{"href":156,"rel":157},"https://doi.org/10.1038/s41586-023-06185-3",[146],[159],{"type":23,"value":156},{"type":17,"tag":25,"props":161,"children":162},{},[163,165],{"type":23,"value":164},"[3]Ravuri, S., Lenc, K., Willson, M. et al. Skilful precipitation nowcasting using deep generative models of radar. Nature 597, 672–677 (2021). ",{"type":17,"tag":142,"props":166,"children":169},{"href":167,"rel":168},"https://doi.org/10.1038/s41586-021-03854-z",[146],[170],{"type":23,"value":167},{"type":17,"tag":25,"props":172,"children":173},{},[174,176],{"type":23,"value":175},"[4]Zhang, Y., Long, M., Chen, K. et al. Skilful nowcasting of extreme precipitation with NowcastNet. Nature 619, 526–532 (2023). ",{"type":17,"tag":142,"props":177,"children":180},{"href":178,"rel":179},"https://doi.org/10.1038/s41586-023-06184-4",[146],[181],{"type":23,"value":178},{"type":17,"tag":25,"props":183,"children":184},{},[185,187],{"type":23,"value":186},"[5]Xiong, W., Xiang, Y., Huang, X. et al. AI-GOMS: Large AI-Driven Global Ocean Modeling System. ",{"type":17,"tag":142,"props":188,"children":191},{"href":189,"rel":190},"https://arxiv.org/abs/2308.03152",[146],[192],{"type":23,"value":189},{"title":7,"searchDepth":194,"depth":194,"links":195},4,[],"markdown","content:technology-blogs:en:2997.md","content","technology-blogs/en/2997.md","technology-blogs/en/2997","md",1776506108913]