Climate Change AI newsletter

Recent and Upcoming Events

IcebreakerOne is hosting two webinars to spark community discussion around a potential Energy Data Ecosystem,

The UN-affiliated Technology Needs Assessment project is hosting a webinar on June 30 focusing on climate technologies in Latin America and the Caribbean.

Registration is open for the IEEE Power and Energy Society General Meeting. This year’s theme specifically highlights the role of data, machine learning, and electric transportation. The conference will be held August 3 - 6.

Registration is open for emPOWER20, which will be held from August 26 - 28. The event will host talks and discussion about how to build a more just, carbon-free, and resilient energy ecosystem.

Recordings are available from the Planet, Smart Cities, and Space session at CogX2020. Tune in to learn how AI and emerging technologies can help make cities and transportation more sustainable.

Calls for Submissions

Due to disruptions resulting from COVID-19, the paper submission deadline for Climate Informatics 2020 has been extended to June 30. The conference will be held September 23 - 25.

Abstract submissions are open for the AGU Fall Meeting 2020, which will be held December 7 - 11. The deadline for abstracts is July 29.

Funding & Entrepreneurship

Applications are open for Rockstart Energy’s startup accelerator, the Energy Program. Selected teams will participate in periodic Deep Dives over the course of six months. Applications due July 26.

Vinnova and Formas, the innovation and research agencies of Sweden, are calling for proposals on AI in the Service of Climate. Proposed projects should support Sweden’s goals of either (1) being net carbon zero by 2045 or (2) adapting to climate change. Proposals due August 25.

AI4Cities' open market consultation phase is ongoing, and webinars from Helsinki, Stavanger, and Copenhagen are now viewable online. This is the first phase of a procurement process on urban energy and mobility solutions.

Data & Challenges

BigEarthNet is now available via Radiant MLHub. The dataset consists of nearly 600K image patches from the European Space Agency’s Sentinel 2 satellite, along with corresponding landcover annotations.

Omdena is hosting a two month solar energy challenge in partnership with SolarAI and ENGIE Factory. Accepted participants will train models to recognize of rooftop features that, if properly catalogued, could facilitate the wider adoption of solar panels. Project starts July 14.

Figures from this month’s readings. A) Feng et al. (2020) show that by measuring consensus between solar panels, it’s possible to detect faults. B) Engram et al. (2020) discover that methane ebullition is visible from lake ice textures. C) Zhang et al. (2020) apply new spectroscopic techniques to better estimate battery capacity. D) Weber et al. (2020) use machine learning to provide surrogate models for precipitation. E) Raghu and Schmidt (2020) conceptualise the scientific deep learning workflow as a multistage, iterative process.

Readings

SunDown: Model-driven Per-Panel Solar Anomaly Detection for Residential Arrays
Menghong Feng, Noman Bashir, Prashant Shenoy, David Irwin, Beka Kosanovic
Summary: Undetected faults in residential solar panel installations lead to unnecessary production loss, but the usual fault-detecting sensors are expensive. This study proposes an approach to fault detection that requires no new hardware. A half-sibling regression is used to formalize the idea that, if a panel is in need of repair, its production will be noticeably lower than its neighbors’. The approach is extended to handle simultaneously failing panels, and the system is validated using field experiments. Discussion.

Remote Sensing Northern Lake Methane Ebullition
Melanie Engram, Katey Walter Anthony, Torsten Sachs, Katrin Kohnert, Andrei Serafimovich, Guido Grosse, Franz Meyer
Summary: Methane is a potent greenhouse gas, and it is important to quantify the amount of methane ebullition (bubbling) out of arctic lakes, especially in light of rapid shifts in climate at northern latitudes. Previously, quantification relied on upscaling of measurements taken at individual lakes. As an alternative, the authors propose the analysis of remote sensing imagery -- the degree of ebullition can be recognized from textures appearing on lake ice. They train models based on emissions recorded at ground stations and evaluate using independent aircraft sensor measurements. Discussion.

Identifying Degradation Patterns of Lithium Ion Batteries from Impedance Spectroscopy using Machine Learning
Yunwei Zhang, Qiaochu Tang, Yao Zhang, Jiabin Wang, Ulrich Stimming, and Alpha A. Lee
Summary: To facilitate wider adoption of lithium ion batteries, it’s important to be able to accurately estimate various properties of their health and degradation. By collecting data on battery degradation in controlled conditions, this study shows how a new sensing technique -- electrochemical impedance spectroscopy (EIS) -- can give better estimates than traditionally used features. One challenge of EIS data is high-dimensionality; the authors address this by using an automatic relevance detection prior in a gaussian process regression model. Code and data are publicly available. Discussion.

Deep Learning for Creating Surrogate Models of Precipitation in Earth System Models
Theodore Weber, Austin Corotan, Brian Hutchinson, Ben Kravitz, and Robert Link
Summary: The development of fast and accurate surrogates for Earth Systems Models could streamline climate science and adaptation planning workflows, for example, enabling the rapid exploration of a range of climate scenarios. The authors evaluate surrogate models based on convolutional neural networks, training and testing using 140 years worth of monthly precipitation data. They find that a deep residual network with scheduled sampling outperforms both traditional surrogates and naive deep learning approaches. Code is publicly available. Discussion.

A Survey of Deep Learning for Scientific Discovery
Maithra Raghu and Eric Schmidt
Summary: This review provides an accessible introduction to the use of deep learning in scientific problem solving. Relevant methods and workflows are described, interspersed with many links to relevant packages, tutorials, and pretrained models. Technical but concise introductions to methods for image, sequential, and graph structured data are given. Finally, the authors highlight emerging challenges within the domain, including working with limited data and the central role of scientific understanding. Discussion.