Research talks: Research partners on innovation for carbon neutralization
\n0:00 Session overview
\nSpeaker: Beibei Shi, Senior Research Program Manager, Microsoft Research Asia<\/p>\n
3:52 Mimicking atmospheric processes of CO2 with a physical-informed deep neural network
\nSpeaker: Jia Xing, Associate Professor, Tsinghua University<\/p>\n
It\u2019s crucial to reduce CO2 emissions worldwide to mitigate and avoid the risks stemming from global climate change. Observations from satellite images can directly measure the global CO2 concentration at a high resolution, and they enable us to track the progress of CO2 controls globally. However, ambient CO2 concentration is influenced by a variety of atmospheric processes involving both anthropogenic and natural sources. This creates a challenge in attributing observed CO2 concentrations to CO2 emissions, particularly for megacities undergoing a change of anthropogenic activities in a relatively short time. Numerical models were mostly used in previous studies to mimic all atmospheric processes involved, while its heavy computational burden required us to compromise between spatial resolution and coverage. In this talk, we propose novel deep neural networks (DNNs) to capture the nonlinear relationship between ambient CO2 concentration and CO2 emissions. Our results suggest that the newly developed DNN-CO2 can well reproduce the atmospheric numerical model in predicting the relationship between ambient CO2 concentration and CO2 emission input. The DNN-CO2 can further be used to infer dynamic CO2 emissions based on observed measurement, which can greatly assist local agencies in evaluating the control\u2019s effectiveness and designing a more effective control strategy for mitigating CO2.<\/p>\n
17:47 Real-time monitoring of global CO2 emissions and the negative carbon computing
\nSpeaker: Zhu Liu, Associate Professor, Tsinghua University<\/p>\n
The diurnal cycle of CO2 emissions from fossil fuel combustion and cement production reflect seasonality, weather conditions, working days, and more recently the impact of the COVID-19 pandemic. In this research talk, we discuss how we\u2019re able to provide, for the first time, a daily CO2 emission dataset for all of 2020 calculated from inventory and near real-time activity data for power generation (29 countries), industry (73 countries), road transportation (406 cities), aviation and maritime transportation, and residential fuel use sectors (estimated for 206 countries). We show from detailed estimates of the data that the global reduction was 6.3% (-2,232 Mt CO2) compared with 2019 levels, while a larger global reduction was 6.5% compared with our baseline simulations by removing the effect of historical trends. This decrease is nearly six times larger than the annual emission drop at the peak of the 2008\u20132009 global financial crisis. These declines are significant but will be quickly overtaken by new emissions, unless we use the COVID-19 pandemic as a breakpoint with our fossil-fuel trajectory, notably through policies that use the COVID-19 recovery as an opportunity to make national energy and development plans more sustainable.<\/p>\n
34:43 Forest fire prediction in the developing world: The power of machine learning
\nSpeaker: Kuldeep S. Meel, Presidential Young Professor, National University of Singapore<\/p>\n
Deforestation and climate change have dramatically increased the number of forest fires across the globe. In Southeast Asia, Indonesia has been most affected by tropical peatland forest fires. These fires have a significant impact on the climate, which results in extensive health, social, and economic problems in societies. To add to this problem, existing forest fire prediction systems are based on handcrafted tools and require installing and maintaining expensive instruments on the ground, which is challenging for developing countries such as Indonesia. In this talk, I will discuss a cost-effective machine learning-based approach that uses remote sensing data to predict forest fires in Indonesia.<\/p>\n
47:05 Offshore CO2 storage in the form of gas hydrate
\nSpeaker: Toru Sato, Professor, University of Tokyo<\/p>\n
Carbon capture and storage (CCS) is a promising technique for reducing significant amount of CO2. Gas hydrate has ice-like structures formed by the enclosure of gas molecules by water molecule cages. CO2 hydrate can be formed in high pressure and low temperature conditions, such as in sub-seabed shallow formations in the deep sea. To use gas hydrate positively, an approach has been proposed: CO2 is injected below the depth at which CO2 hydrate forms. To estimate CO2 hydrate\u2019s sealing effect against the upward leakage of liquid CO2, reservoir-scale numerical simulations in realistic geological formations are necessary. However, the conventional permeability model used in the simulation has poor accuracy for the permeability reduction caused by hydrate formation. To evaluate the permeability reduction, a pore-scale numerical method is a powerful tool because it can investigate microscopic hydrate distribution that controls the permeability. Consequently, a multi-scale simulation consisting of the reservoir-scale and the pore-scale simulations is strongly expected. However, the pore-scale simulation is time-consuming, and it\u2019s nearly impossible in today\u2019s computing facilities to couple the two simulators. In this talk, we\u2019ll discuss our research goals of developing a multi-scale simulator based on a reservoir-scale simulator using neural networks trained with large numbers of data resulting from pore-scale simulations.<\/p>\n
1:00:29 Low carbon transformation pathway for China\u2019s coal-powered plants
\nSpeaker: Xinran Wei, Researcher, Microsoft Research Asia<\/p>\n
In China, about 4,600 coal-powered plants need to make a retrofitting decision between 2025 and 2060 to realize the country\u2019s carbon-neutrality goal. However, the heterogeneity of the power plants requires that these decisions be made plant-by-plant to achieve their respective emission reduction target in the most cost-effective way. This work has never been done because it requires a great deal of data to make plant-level decisions, and the dynamic decision-making process is so complex that it requires huge computing power and innovative algorithms. In this talk, we discuss how the computing resources and advanced machine learning algorithms of Microsoft Research Asia can greatly facilitate finding the optimal solution for this problem. The results of this project and the suggested plant-level retrofitting pathway has great academic value and will provide practical guidance in creating the carbon neutrality roadmap in China.<\/p>\n
1:11:43 The next generation of GNN: From molecular modeling to carbon removal
\nSpeaker: Shuxing Zheng, Senior Researcher, Microsoft Research Asia<\/p>\n
In this talk, we\u2019ll introduce Graphormer, the next generation graph neural network (GNN), which won the first prize in KDD Cup 2021 for the OGB-LSC challenge (quantum chemistry track). We\u2019ll briefly introduce the background of carbon removal material discovery. We\u2019ll then discuss the Graphormer, a powerful GNN built on standard transformer architecture and equipped with three simple and effective structural encodings. We\u2019ll also explore the potential impacts of the many graph-related applications of sustainability.<\/p>\n