Toward a Deeper Understanding of Our Climate System Through Data Science | Emily Gordon

Summary

在这段视频中，斯坦福大学的数据科学家Emily Gordon讨论了气候变化这一关键的可持续性领域。她展示了全球海表面温度的年度平均值，并讨论了自1950年以来全球变暖的趋势。Gordon强调了气候模型在理解气候系统和预测温室气体排放影响方面的重要性。她介绍了气候模型如何帮助科学家模拟不同的气候情景，并预测未来气候变化对地区性极端天气事件的影响。此外，她还提到了使用数据科学方法来分析大量气候模型数据，以及如何通过改进模型本身，例如利用神经网络和人工智能技术，来提高气候模型的准确性和效率。

Takeaways

• 🌡️ 全球海表面温度自1950年代以来呈现上升趋势，显示了全球变暖的现象。
• 📊 确定温室气体排放对气候变暖的贡献是一个复杂的问题，需要依赖气候模型来辅助理解。
• 🌐 气候模型是理解气候系统的强大工具，它们通过不同的假设和参数化方法模拟地球系统。
• 🔮 气候模型可以进行各种实验，比如模拟地球升温、降温或模拟火山爆发等极端事件的影响。
• 📉 气候模型预测显示，二氧化碳浓度加倍可能导致2到4摄氏度的全球变暖。
• 📈 气候模型的预测能力在过去几十年中得到了验证，1979年的预测与当前的模型结果相符合。
• 🌍 气候模型可以预测不同气候敏感度下的全球变暖情况，并预测未来可能超过1.5摄氏度的阈值。
• 🌡️ 气候模型还可以用来预测区域尺度上的温度变化，如热浪和极端降雨事件的增加。
• 🔬 构建气候模型需要将地球系统划分为网格盒，并使用参数化方法来处理小于网格盒尺度的物理过程。
• 💾 气候模型产生的数据量巨大，需要先进的数据科学方法来分析和理解这些数据。
• 🌐 未来的工作包括使用现有数据源进行气候预测、训练AI模型快速查询气候影响、以及改进气候模型本身。

Q & A

全球海表面温度的变化趋势是什么？

从1950年代开始，全球海表面温度呈现上升趋势，这表明了全球变暖的现象。

温室气体排放对全球变暖的贡献有多大？

温室气体排放是导致全球变暖的主要原因之一，但实际观测到的变暖幅度略低于气候模型预测的仅由温室气体排放引起的变暖。

气候模型是如何帮助我们理解气候变化的？

气候模型通过模拟地球系统的不同假设和过程，使我们能够进行实验，比如增加温室气体浓度，观察地球系统的响应，从而帮助我们理解气候变化。

气候模型的预测准确性如何？

气候模型的预测在历史上显示出相当的准确性。例如，1979年的预测与后来的观测结果相比，预测的变暖范围与实际观测相当接近。

气候模型中的“敏感性”是什么意思？

气候模型的敏感性指的是模型对二氧化碳浓度变化的反应程度，高敏感性模型预测的变暖幅度更大。

气候模型如何帮助我们理解区域气候变化的影响？

气候模型可以链接到区域尺度，帮助我们理解不同水平的全球变暖对热浪和极端降雨事件等区域气候影响。

构建气候模型需要考虑哪些因素？

构建气候模型需要考虑地球系统的网格划分、子网格尺度的过程（如云和地形）、辐射方案的准确性，以及通过参数化来解决小尺度过程。

气候模型产生的数据量有多大？

最新的气候模型实验产生了约20拍字节（petabytes）的数据，这显示了气候模型产生的数据量是巨大的。

如何使用现有的气候模型数据进行科学研究？

可以使用现有的气候模型数据来训练仿真器和AI，快速有效地查询特定年份或变量的区域气候变化影响。

气候模型中存在哪些挑战和改进方向？

气候模型中的挑战包括处理大量数据、理解内部气候过程、减少系统性偏差以及提高模型的可信度和稳健性。改进方向包括使用AI和神经网络来改进子网格尺度的参数化。

气候模型如何帮助我们应对未来的气候变化？

气候模型不仅可以帮助我们预测未来的气候变化，还可以通过数据驱动的方法进行科学发现，提高模型的准确性和效率。

Outlines

00:00

🌡️ 全球变暖与气候模型

本段讨论了气候变化的关键领域之一——全球变暖。主讲人Emily Gordon介绍了全球海表面温度的年度平均值，并将其与1850至1900年间的平均值进行了比较。自1950年以来，我们观察到了明显的全球变暖现象。讨论了温室气体排放对温度上升的贡献，并指出这是一个复杂的问题，因为观测时间尺度与我们拥有的数据量相当。气候模型在这里发挥了重要作用，它们可以帮助我们理解地球系统的不同方面，通过不同的假设和参数化方法来模拟地球系统。主讲人还提到了气候模型如何帮助我们预测未来可能的气候变化，以及如何使用这些模型来评估不同温室气体排放情景下的气候敏感性。

05:01

🌍 气候模型的构建与挑战

第二段深入介绍了气候模型的构建过程，包括将地球系统划分为网格盒子来求解运动方程，并在不同时间步长之间传递信息。由于存在比网格盒子更小尺度的气候现象，如云层和地形，因此需要通过参数化过程来近似这些小尺度现象。主讲人提到了模型实验中产生的大量数据，大约有20拍字节（petabytes）的数据可用。此外，还讨论了气候模型中的系统性偏差，以及如何确保模型能够提供给利益相关者准确可靠的信息。主讲人还展望了未来，包括使用现有数据源进行气候变化预测、训练AI模型以快速查询气候影响、进行下尺度化处理以将大尺度气候信息应用于小尺度分析，以及使用神经网络和AI改进气候模型的子网格参数化。

10:04

📈 数据科学在气候模型中的应用与未来展望

最后一段强调了数据科学在气候模型分析中的应用，并展望了未来的发展。主讲人提到了如何使用数据驱动的方法来发现新的科学知识，并且强调了跨学科合作的重要性，即结合数据科学和领域知识来解决气候问题。此外，还提到了使用AI和机器学习技术来改进气候模型，使其运行更快、更准确。最后，主讲人以2月4日加利福尼亚州的大气河流事件的卫星图像作为结尾，强调了数据在理解气候变化中的重要性。

Keywords

💡气候变化

气候变化是指地球气候系统长期变化的趋势，通常与全球变暖联系在一起。在视频中，气候变化是主要讨论的主题，特别是它与温室气体排放的关系。例如，视频提到自1950年以来观测到的全球平均海表面温度的升高，这被归因于气候变化。

💡数据科学

数据科学是一种利用数据进行分析和解释的跨学科领域，它在视频中被用来理解和解决气候变化问题。例如，数据科学被用于补充观测数据不足，通过气候模型来预测未来的气候变化。

💡气候模型

气候模型是模拟地球气候系统行为的数学模型。视频中提到，由于观测数据的局限性，气候模型成为理解气候变化的关键工具。模型可以模拟不同的气候情景，如温室气体排放增加或减少，以及它们对地球气候的影响。

💡温室气体排放

温室气体排放是指人类活动产生的气体，如二氧化碳，这些气体能够吸收和重新辐射热量，导致全球变暖。视频中讨论了温室气体排放对全球平均海表面温度的影响，并用模型预测了其潜在的气候变化效应。

💡气溶胶排放

气溶胶是悬浮在空气中的微小颗粒，它们可以反射太阳光，从而对气候产生冷却效应。视频中提到气溶胶排放与温室气体排放相反，有助于抵消一部分由温室气体引起的全球变暖。

💡碳排放

碳排放通常指的是二氧化碳的排放，这是最主要的温室气体之一。视频中提到了碳排放对气候变化的影响，并讨论了如何通过气候模型来预测在二氧化碳浓度加倍情况下的全球变暖程度。

💡气候敏感性

气候敏感性是指气候系统对特定强迫因素（如温室气体浓度变化）的响应程度。视频中提到了不同气候模型对二氧化碳的敏感性，并预测了在不同敏感性下的未来气候变化。

💡极端气候事件

极端气候事件指的是异常的气候现象，如热浪、极端降雨等。视频中提到了随着全球变暖，这些事件的频率和强度可能会增加，并展示了气候模型如何预测这些极端事件的变化。

💡参数化

参数化是一种在气候模型中处理小尺度过程的方法，如云的形成和辐射传输。由于这些过程太复杂或尺度太小，无法直接在模型中解决，因此需要通过参数化来近似它们的影响。视频中解释了参数化在构建气候模型中的重要性。

💡气候模型的偏差

气候模型的偏差指的是模型预测与实际观测数据之间的差异。视频中强调了理解这些偏差的重要性，以确保模型的可靠性和预测的准确性。

💡数据驱动方法

数据驱动方法是指利用大量数据进行分析和预测的方法。视频中提到了使用数据驱动方法来改进气候模型，提高预测的准确性，并强调了这种方法在科学发现中的潜力。

💡人工智能

人工智能（AI）是指使计算机系统模拟人类智能的技术。视频中提到了利用AI来提高气候模型的效率和准确性，例如通过训练模拟器来快速查询特定气候变量的影响。

💡下尺度化

下尺度化是一种将大尺度气候模型的数据转换为小尺度的方法，以便更好地理解区域气候影响。视频中提到了下尺度化在将气候模型的预测应用于地方或区域层面的重要性。

💡科学发现

科学发现指的是通过研究和分析获得的新知识和理解。视频中提到数据驱动方法不仅有助于分析现有数据，还有可能揭示新的科学现象或理论，从而推动科学的进步。

Highlights

海面温度数据显示自1950年代以来全球变暖现象。

气候模型是理解气候系统的关键工具，通过不同假设和过程模拟地球系统。

气候模型可以模拟不同的气候情景，如增温、冷却或火山爆发等。

通过气候模型的实验，可以估计二氧化碳加倍对全球变暖的影响。

1979年的气候模型预测在二氧化碳加倍情况下，全球变暖幅度在2到3.5度之间。

最新的IPCC报告预测，二氧化碳加倍可能导致2.5到4度的全球变暖。

气候模型的多样性允许我们评估不同模型对气候强迫的响应。

气候模型预测显示，我们可能在未来10年内超过1.5度的全球变暖阈值。

气候模型数据可以链接到区域尺度的地表影响，如热浪和极端降雨事件。

气候模型的构建需要将地球系统划分为网格盒，并解决运动方程。

参数化是处理小于网格盒尺度的气候过程的关键技术。

气候模型产生的数据量巨大，需要有效的数据科学方法进行分析。

气候模型的系统性偏差需要被理解和校准，以提供可靠的信息。

数据科学方法不仅可以预测气候变化，还可以训练AI进行快速查询。

降尺度是将大尺度气候信息转化为小尺度影响的关键领域。

负责任和可信赖的AI方法对于气候模型的数据分析至关重要。

数据驱动的方法为科学发现提供了新的机会。

使用神经网络和AI改进气候模型的子网格参数化，提高模型的准确性和效率。

气候模型和数据科学方法的结合为理解气候变化提供了新的视角。

气候模型和观测数据的结合为未来气候模型的改进提供了方向。

Transcripts

[Music]

next up in this block we'll hear about

one of the critical areas of

sustainability and that's climate change

and for that we have Stanford data

science postto Emily Gordon Emily please

join

us so um I'm really excited to share

some of the um challenges that we face

um understanding climate change and sort

of bringing data science um into the

process so here I'm showing a plot of

annual mean Global sea surface

temperatures and I've plotted this as a

departure from the 1850 to about 1900

mean and you can see from about 1950 uh

1950 onwards we have our uh Global mean

uh global

warming so we might ask uh you know how

much uh you know warming is attributable

to greenhouse gas emissions and it's

actually a really hard problem to figure

out what the amount of observations that

we have because the time scales that

we're interested on are about the same

as the amount of data we have from our

observational methods so we turn to

climate models so this black line I'm

showing here is the same as the Green

Line in the other figure it's our

observed Trend in Sea surface

temperatures um but in the Gray Line I'm

showing the uh from a climate modeling

experiment what we would actually expect

how much warming we expect from

greenhouse gas emissions alone so we see

we have slightly more warming than we

actually have currently observed and

that's because of this blue line which

is aerosol emissions so this plot to me

at the sort of takehome here is that we

don't have enough observations to

understand completely what's going on in

our climate system and so we have to

supplement this with uh climate model

data to really pick about pick out what

our what is going on in our

system so climate models are an

incredibly powerful tool for

understanding our climate we can think

of them as all these different models as

sort of different ways of uh uh uh

thinking about the Earth system we can

make different assumptions we can

include different processes we can

parameterize them in different ways so

we can get all these different

realizations of our Earth system and we

can perform experiments on them we can

do things like warm up the planet we can

cool it down we can hit it with big

shocks like volcanic eruptions we can

remove components like the land and see

uh how the fluid flows without you know

uh topography or we can do smaller scale

things like brighten clouds and see how

that affects um the climate

system but another part of the power of

these tools is that because we have all

of these different models with these

different assumptions

we can perform the same experiment on

all of our models and see how they

respond and and and sort of say

something about our our own planets

response to that forcing so for example

if we were to double the amount of

carbon dioxide in the atmosphere we

might want to look at how much warming

we would get across all of our different

climate models and then build a

distribution from these models and we

can put a sort of an uncertainty

estimation on how much warming we might

expect um in the

future and in fact this is a pretty

common experiment to perform so this was

first performed in about

1967 um and then I've taken this quote

from the Chie report in 1979 who looked

at um using simple climate models they

found that we would expect between 2 to

3.5 degrees warming under a doubling of

carbon

dioxide and this experiment has still is

still being run to this day it was um

included in the most recent ipcc

International uh intergovernmental panel

on climate change report and this is the

result from the most like from the most

recent report where they find the likely

range of warming is between 2.5 and 4°

so actually we were doing really well

with our climate models back um in

1979 so we can take all of our climate

models and we can rank them by how

sensitive they are to carbon

dioxide and then we can project them

forward under different scenarios of

climate change so uh for you know more

High sensitivity uh climate models we

would have expected to have passed this

1.5 degree threshold which is in the

sort of dark this orangey color um you

know a couple years ago so maybe we're

not quite so sensitive to to carbon

dioxide as these really high sensitivity

models but we may be looking at Crossing

1.5 in the next sort of 10 years or

so and then because these climate models

are are run you know globally we can

then link the sort of warming that that

we get from these different models to

the surface impacts on Regional scales

so this top uh figure here is showing

the increases in heat events under

different levels of warming and then the

bottom is showing the increases in uh

extreme rainfall events under the

different level of

warming and I also want to just

reinforce from just these three figures

that I've shown the sheer amount of data

that's gone into making them so there um

from that probability distribution

there'd be about at least 50 climate

models run between 200 to 1,000 years of

data from Those runs and then on top of

that those scenarios there were five

scenarios with three members each so

that's another 300 years or so uh so

there's just so so much data that we're

churning out from our climate models

now so how do we build a climate model

so we take our Earth system and we need

to divide it up into our grid boxes

where we're going to solve our equations

of motion and sort of pass the

information between the boxes at

different time

steps but we have things that are going

on on smaller than our grid Box level uh

we have things like clouds we have to

understand our bottom topography so

whether we're over land over the ocean

you know over mountains and we also have

to get the radiation scheme correct and

so we do this uh by a process called

parameterization so we uh make sort of

empirical functions to solve what's

going on in these um subg good scale

levels and this has to be done for all

of the different processes so I'm

showing us the algorithm for one type of

cloud process within a within a grid

box so from our most recent round of

modeling experiments we have about 20

pedabytes of data available we're not

just interested in our response to risk

forcing which is what I've shown so far

we're also interested in um variability

and predictability so understanding

actually our internal climate processes

how they project onto clim climate

change and then um cause some of these

more extreme events that we've seen and

another really important part is the

systematic biases and our climate models

making sure we understand the processes

that we are realizing and the processes

that we are not so if we're providing

information to stakeholders that we know

that we have good calibrated

uncertainty and then this is a figure

that projects forward the amount of data

that we're going to be having in the

next uh sort of 10 years or so and and

we just this this data challenge is just

going to become uh more and more St

um and and there's so much room for for

bringing in good data science methods

for analyzing all of this

data so what are we doing um in our

future so first of all I've already

shown this using our existing data

sources is uh making projections of

future climate change uh using the

existing modeling runs to uh sort of

pull out these different um impacts of

climate

change this uh this one's slightly

different this is looking at using all

of the data we have to train uh

emulators to train AI so that you can

quickly and and efficiently query you

know um you might be interested in in a

carbon dioxide forcing a certain year a

certain variable and pull out what the

regional impact of climate change um

that you're interested

in another thing is a process called

downscaling so I mentioned that climate

models are run at pretty CSE resolution

but you know um we uh we we we feel the

climate impacts on our on our small grid

boxes you know um in in our day-to-day

lives and so there's a lot of uh effort

now in in bringing our big large scale

climate information to the smaller

scales and this is a really exciting

area because we are linking our observed

relationships between the large scale we

can go out and do field campaigns we can

understand how the large scale

variability um impacts the smaller

scales to sort of bring this large scale

understanding of climate models to um

the smaller scale impact

analysis and then finally um sort of

ethical responsible trustworthy Ai and

data driven methods if all of our

information is coming from or most of

our information is coming from

datadriven analysis on climate models we

want to make sure that any information

that we're providing is absolutely

trustworthy and

robust the other thing I think that's

actually really exciting here is this

actually has allowed for scientific

discovery so if we're using our data

driven methods to understand um to uh to

a certain response to climate change and

we find something that's not expected

well there could be two reasons for this

one you've done something wrong oh no

that's bad but you you do all these

checks and you go okay actually know you

know everything is still meshing with my

physical understanding of the system so

so this is an opportunity to discover

new

science the other thing that I think

that I really want to un underscore here

is that not only is is this a data

science effort but this is a physical

science eff and this is um I think

bringing together sort of this

interdisciplinary theme of of of this

meeting that we that we need to be

working together with our data science

understanding and with our domain um

knowledge and then I wanted to finish

with this final thing that's um uh

become sort of a really a really great

Avenue for future Improvement is not

only like thinking of our climate models

as you know these things that are over

here existing tuning out data but how

can we improve the models themselves so

I talked before about how we have these

sub grid scale

parameterizations what about if we start

replacing them with with neural networks

with AI we can use again our

observational data of the relationships

between um between you know different

processes and and and really quickly

build um empirical relationships to make

um to make uh our climate models better

and and run more quickly and and and I

saw a talk about this a couple of days

ago and it's just so so impressive what

what people are doing with

this so I'm finishing here with a

satellite Loop of an atmospheric River

Event that was over California on the

4th of February if people remember it um

and and I I guess I want to also

reinforce that uh you know not only do

we have these beautiful pictures but

this data is just so so important to us

and our understanding of climate

change thank

[Applause]

you