Past Projects

Here, I list several research projects I have completed in the past. So you can read into more details about them compared to the short version in my CV.

Forecasting tropical cyclone trajectories and predicting their intensity is critical for protecting people and property in due time. Currently, operational forecast models consist of three types: dynamical models, which are based on solving physical equations; statistical models, which utilise historical observations and meteorological features; and a mixture of both. The most accurate models are often sophisticated dynamical models. On the other hand, statistical forecasting models have shown recent progress in modeling such time-series but their full potential is yet to be realized. In particular, there is room for state-of-the-art machine learning methods to be tested in the context of hurricane forecasting. In this work, we adapt state-of-the-art Computer Vision and tree-based methods for hurricane track and intensity forecasting. We show our models are competitive with current operational forecasts and can even outperform the National Hurricane Center’s official forecast when included in a consensus model.

To learn more, you can find our paper on Arxiv here.

This is my undergraduate research project. I was intrigued by the sense of surprise after the 2016 UK-EU Referendum result, and decided to seek for a mathematical explanation for this volatile opinion shift. I used an agent-based modelling framework, and applied statistics mechanics to describe the dynamics of opinion influence among individuals.

The model assumes that some of the individuals are more stubborn than others regarding political viewpoints, that they are unlikely to change their mind. Therefore, we separate our agents into: open-minded and stubborn ones, each holding initially an opinion: A or B. We let the group interact according to some rules and study how stubborn-agents alter the opinion formation dynamics.

In particular, at every time step, we perform the following three steps:

  1. Randomly distribute agents into small subgroups of size r.
  2. Open-minded agents update to local majority within the subgroup, stubborn agents unchanged
  3. Reshuffle

This model aims to reflect the information exchange behaviour and people's different degree of flexibility when influenced by others. The stubborn agents essentially function as "influencers" because they only contrbute to influencing others but do not ever change their mind. Mathematically, if we assume the group size is large enough, we can describe the support for opinion A (i.e. number of agents supporting A) at each time step using statistical mechanics.

The mathematical model allows one to study how does the initial composition of agents (such as the level of stubborn agents, and the initial opinion allocatin) lead to different opinion evolvement, and hence to different voting outcomes. A surprising result is that in some cases, stubborn agents can entirely determine the direction of opinion formation. In addition, reducing the subgroup size can enhance such "stubborn effect", suggesting the internet contribute to opinion manipulation in our current society, implying the danger of Populism. As an application, this model can be applied to forecast election outcome using polling data because it can capture the trend of the opinion formation.

To learn more, you can find my my report here and a presentation here.

I completed a summer research internship at the MIT Media Lab in 2018. The project aims to develop a complex systems modelling approach to study the coastal mangrove ecosystem in Rio de Janeiro, which is under two-sided threats from landward urban expansion pressure and seaward sea-level rise pressure.

Firstly, employed knowledge from biology and urban planning to develop a logic map, represenging the complex dynamics from various agents: urban units, agriculture units, mangrove units, the sea, transportation network. Then, we simulated this multi-agent systems model computationally using landu use data from Rio de Janeiro. Illustrated below is a logic map and a summary of computational steps to simulate the growth (or loss) of mangrove forests. This project serves as a prototype to develop computational tools for urban planning, which can intake both environmental and social factors. We collaborated with mangrove biologists and urban planning officials from Rio.

To learn more, you can read my report here and a visualisation here. Underlying logic map Computational steps to simulate mangrove growth