Exploring Deep Learning for Earth System Forecasting

PhD Thesis Proposal Defence


Title: "Exploring Deep Learning for Earth System Forecasting"

by

Mr. Zhihan GAO


Abstract:

Conventionally, Earth system (e.g., weather and climate) forecasting relies on 
numerical simulation with complex physical models and hence is both expensive 
in computation and demanding on domain expertise. With the explosive growth of 
spatiotemporal Earth observation data in the past decade, data-driven models 
that apply Deep Learning (DL) are demonstrating impressive potential for 
various Earth system forecasting tasks, including precipitation nowcasting, 
ENSO forecasting and Earth surface forecasting. However, previous DL approaches 
for Earth system forecasting typically relied on the combination of Recurrent 
Neural Networks (RNN) and Convolutional Neural Networks (CNN). These 
approaches, however, faced challenges such as incorrect inductive biases, poor 
scalability, and failing to handle uncertainty. This thesis aims to explore how 
recent advancements in Deep Learning (DL) techniques can overcome these 
challenges, significantly benefiting Earth system forecasting.

First, we propose Earthformer, a space-time Transformer for Earth system 
forecasting. Earthformer is based on a generic, flexible and efficient 
space-time attention block, named Cuboid Attention. The idea is to decompose 
the data into cuboids and apply cuboid-level self-attention in parallel. These 
cuboids are further connected with a collection of global vectors. Earthformer 
achieves state-of-the- art performance on two synthetic datasets MovingMNIST 
and N-body MNIST, and two real-world benchmarks about precipitation nowcasting 
and El Niño/Southern Oscillation (ENSO).

Second, we propose PreDiff, a conditional latent diffusion model for 
precipitation nowcasting, along with a generic two-stage pipeline for 
probabilistic precipitation nowcasting: 1) developing a purely data-driven 
model capable of probabilistic forecasts. 2) incorporating an explicit 
knowledge alignment mechanism to align forecasts with domain-specific physical 
constraints. Experiments demonstrate the effectiveness of PreDiff in handling 
uncertainty, incorporating domain-specific prior knowledge, and generating 
forecasts that exhibit high operational utility.


Date:                   Monday, 6 May 2024

Time:                   12:30pm - 2:30pm

Venue:                  Room 5501
                        Lifts 25/26

Committee Members:      Prof. Dit-Yan Yeung (Supervisor)
                        Prof. Albert Chung (Chairperson)
                        Dr. Yangqiu Song
                        Prof. Raymond Wong