I was reading through the whitepaper and was wondering what variables were fed back into the model
\n\n\nOn the other hand, “Online evaluation” has domain-specific meaning in our context,\nand refers to evaluating the performance of the hybrid climate simulator in which many\ninstances of its embedded high resolution physics solver are replaced with copies of a\ntrained ML parameterization that is then allowed to feed back with resolved planetary scale\nclimate dynamics.
\n
https://arxiv.org/abs/2306.08754
\nI wrote a rain simulator over a several hour period but the models could only use: state_t, state_q0001-3, and state_u, state_v (and their corresponding ptend* values) because those were the only variables that could change over time.
\nhttps://www.kaggle.com/zahid1/rain-simulator
How were the additional variables used to feed back into the model? How were columns without dX/dt used to simulate future values?
", "messages": [ { "id": 2941452, "postDate": "2024-07-31T01:30:14.107Z", "content": "I was reading through the whitepaper and was wondering what variables were fed back into the model
\n\n\nOn the other hand, “Online evaluation” has domain-specific meaning in our context,\nand refers to evaluating the performance of the hybrid climate simulator in which many\ninstances of its embedded high resolution physics solver are replaced with copies of a\ntrained ML parameterization that is then allowed to feed back with resolved planetary scale\nclimate dynamics.
\n
https://arxiv.org/abs/2306.08754
\nI wrote a rain simulator over a several hour period but the models could only use: state_t, state_q0001-3, and state_u, state_v (and their corresponding ptend* values) because those were the only variables that could change over time.
\nhttps://www.kaggle.com/zahid1/rain-simulator
How were the additional variables used to feed back into the model? How were columns without dX/dt used to simulate future values?
", "rawMarkdown": "I was reading through the whitepaper and was wondering what variables were fed back into the model\n\n>*On the other hand, “Online evaluation” has domain-specific meaning in our context,\nand refers to evaluating the performance of the hybrid climate simulator in which many\ninstances of its embedded high resolution physics solver are replaced with copies of a\ntrained ML parameterization that is then allowed to feed back with resolved planetary scale\nclimate dynamics.*\n\n[https://arxiv.org/abs/2306.08754](url)\nI wrote a rain simulator over a several hour period but the models could only use: state_t, state_q0001-3, and state_u, state_v (and their corresponding ptend* values) because those were the only variables that could change over time. \n[https://www.kaggle.com/zahid1/rain-simulator](url)\n\nHow were the additional variables used to feed back into the model? How were columns without dX/dt used to simulate future values?", "votes": 1 }, { "id": 2941535, "postDate": "2024-07-31T04:37:06.887Z", "content": "Those additional scalar output features (e.g., precipitation, surface radiation fluxes) are used to drive a land module in the E3SM Earth system model. The land module can in turn influence the atmosphere module through land-atmosphere interaction (primarily through surface fluxes). I'm not very familiar with the land module, but I believe the land module likely also carries some land variables that vary over time, e.g., soil moisture. For the ocean surface, the E3SM Earth system model has a mode to use prescribed sea surface temperature and sea ice coverage. So those scalar output features won't feed back into the ocean module but only land module.
", "rawMarkdown": "Those additional scalar output features (e.g., precipitation, surface radiation fluxes) are used to drive a land module in the E3SM Earth system model. The land module can in turn influence the atmosphere module through land-atmosphere interaction (primarily through surface fluxes). I'm not very familiar with the land module, but I believe the land module likely also carries some land variables that vary over time, e.g., soil moisture. For the ocean surface, the E3SM Earth system model has a mode to use prescribed sea surface temperature and sea ice coverage. So those scalar output features won't feed back into the ocean module but only land module.", "votes": 2 }, { "id": 2942005, "postDate": "2024-07-31T14:22:56.787Z", "content": "As a follow up: Is there an example of how these revised models (the ones derived in the competition) may be used in practice? Or perhaps just a link to the land module (the one you mentioned above).
\nThank you!
", "rawMarkdown": "As a follow up: Is there an example of how these revised models (the ones derived in the competition) may be used in practice? Or perhaps just a link to the land module (the one you mentioned above).\n\nThank you!", "replies": [ { "id": 2942072, "postDate": "2024-07-31T15:20:43.543Z", "content": "Yes, in practice we can put the trained neural net models into E3SM climate simulator and replace a module that NN emulates (i.e., a cloud-resolving model that solves how cloud and storms evolve in each atmosphere column and feed back to the column's temperature, moisture, cloud, wind, precipitation, etc). One research question would be how well can such hybrid simulator reproduce the behavior of the original pure-physics simulator. Here are a few reference code repo below to test the performance of the hybrid simulator. I don't have an immediate reference to the land module document.
\nclimsim-online repo: a containerized workflow to run the climate simulator (E3SM) with the neural nets in it (to replace the original embedded cloud-resolving model).
\nclimsim repo: it documents how to prepare the neural net model (e.g., the ones from this competition) to put them into the E3SM to run hybrid climate simulations.
Those additional scalar output features (e.g., precipitation, surface radiation fluxes) are used to drive a land module in the E3SM Earth system model. The land module can in turn influence the atmosphere module through land-atmosphere interaction (primarily through surface fluxes). I'm not very familiar with the land module, but I believe the land module likely also carries some land variables that vary over time, e.g., soil moisture. For the ocean surface, the E3SM Earth system model has a mode to use prescribed sea surface temperature and sea ice coverage. So those scalar output features won't feed back into the ocean module but only land module.
", "votes": 2, "replies": [] }, { "id": 2942005, "author_name": "zpanj", "author_url": "", "post_date": "2024-07-31T14:22:56.787000", "content": "As a follow up: Is there an example of how these revised models (the ones derived in the competition) may be used in practice? Or perhaps just a link to the land module (the one you mentioned above).
\nThank you!
", "votes": 0, "replies": [ { "id": 2942072, "author_name": "Zeyuan Hu", "author_url": "", "post_date": "2024-07-31T15:20:43.543000", "content": "Yes, in practice we can put the trained neural net models into E3SM climate simulator and replace a module that NN emulates (i.e., a cloud-resolving model that solves how cloud and storms evolve in each atmosphere column and feed back to the column's temperature, moisture, cloud, wind, precipitation, etc). One research question would be how well can such hybrid simulator reproduce the behavior of the original pure-physics simulator. Here are a few reference code repo below to test the performance of the hybrid simulator. I don't have an immediate reference to the land module document.
\nclimsim-online repo: a containerized workflow to run the climate simulator (E3SM) with the neural nets in it (to replace the original embedded cloud-resolving model).
\nclimsim repo: it documents how to prepare the neural net model (e.g., the ones from this competition) to put them into the E3SM to run hybrid climate simulations.