Tackling climate change with machine learning [part 5] – Industry & carbon dioxide removal

No Comments

On 10th of June, 2019, twenty-two AI researchers, including Andrew Ng and Yoshua Bengio, published a paper on how climate change can be tackled with machine learning. I really enjoyed reading it and I am convinced that the paper as well as the climatechange.ai initative, which emerged from it, deserve more attention. For that reason i created a series of blog posts and videos which provide a dense summary, listing many of the proposed solutions and linking research work as well as ongoing projects. In the big picture, all solutions aim to reduce greenhouse gas emissions.

As my contribution to the global #ClimateStrike week from September 20th to 27th, i will post one chapter (video and blog post) on every working day. You can subscribe to our YouTube channel or follow me on Twitter to be notified when new content is published.

This is part five of a six-part series:

  1. Electricity Systems
  2. Transportation
  3. Buildings & Cities
  4. Farms & Forests
  5. Industry & Carbon Dioxide Removal
  6. Datasets & further resources (available from Friday, Sept. 27th)

Industry

Greenhouse gas emissions caused by industry are often hard to eliminate. However, industries collect more data than ever and cloud storage and computing is becoming more affordable. It is estimated that 60 to 70% of industry data is not used, though a challenge is to get access to it since most of it is proprietary.

Machine Learning can help reduce the carbon footprint of industry by

Reducing waste in supply chains

[high leverage]

In many industries, products are regularly overproduced and overstocked. Excess products not only waste resources through their production but also cause greenhouse gas emissions when shipped and stored in climate-controlled warehouses. As an example, the clothing industry sells on average only 60% of their products at full price. However, some brands can sell up to 85% of their stock using just-in-time manufacturing and effective supply chain networks.

ML can help

  • reduce overproducing and overstocking by improving demand forecasting [Paper – Ensemble learning] [Paper – Multi-Target Regression with Rule Ensembles] [Paper – End-to-End-Learning, Probabilistic Model]
  • decrease post-harvest losses by identifying when produce is about to spoil and ruin bigger parts of a shipment (speculation) [Paper]

Potential Pitfall: More accurate demand forecasts in some cases allows for just-in-time manufacturing. If this leads to smaller and faster shipments of goods that lack the energy-efficiency of freight aggregation, the greenhouse gas effect might not improve as planned.

Reducing material by inventing new constructions

[high leverage] [high risk]

Cement and steel production

Cement and steel production cause astonishing 9% of global greenhouse gas emissions. If the cement industry would be a country, it would emit more greenhouse gases than every other country, except China and the US. The major reason for it is that in order to create cement, materials have to be heated to around 1500 C°. 3D printing allows for unusual shapes that use less material but may be impossible to produce through traditional metal-casting or poured concrete.

Generative design algorithms can reduce material use

Generated design with less resources [Grossman et al., 2017]

ML can help

  • develop structural products that require less raw material [Paper – Generative Design]
  • improve simulation of the physical processes of 3D printing [Paper]
  • push materials science for low-carbon concrete / steel, as discussed in the chapter Electricity Systems: Accelerating materials science

Reducing factory energy consumption

[high leverage]

As discussed in chapter Buildings & Cities: Enabling smart buildings, ML can help reduce energy consumption from heating, ventilation and air conditioning, which can of course be applied within factories as well.

ML can help

  • optimize cooling of giant data centers, e.g. at Google [Blog post – Reinforcement Learning] [Paper – Neural Network]
  • detect hydrofluorocarbon (an extremely potent greenhouse gas) leaks with computer vision and deep learning
  • optimally schedule energy-intensive processes such as cement crushing and powder-coating by providing better demand response optimization algorithms [Paper]

Carbon Dioxide Removal

Until now, we have looked at many solutions that lower emissions, but even if no new carbon would be emitted at all, today’s level of carbon in the atmosphere will continue to have negative consequences in the future. The natural removal of CO2 for example by absorption through plants won’t suffice. Therefore, actively removing CO2 from the atmosphere and storing it in some other form, is necessary. The most promising approach is to sequester CO2. There is also a technique called direct air capture where CO2 sorbents are used to directly filter carbon where it is emitted but this technology’s effectiveness is still low.

Sequestering CO2

[high leverage] [long-term] [high risk]

Carbon dioxide removal

Actively sequestering carbon dioxide can be done by capturing it and storing it underground. A Norwegian oil company has successfully sequestered CO2 from an offshore natural gas field underground for more than twenty years. There are a few improvements that can be achieved with ML.

ML can help

  • identify and characterize potential storage locations [Paper – Neural Networks]
  • monitor and maintain active sequestration sites and help with simulations [Paper – Deep Convolutional Encoder-Decoder Neural Network]
  • monitor potential CO2 leaks from wells [Paper]

Credits

Many thanks to all researchers of the paper:

David Rolnick, Andrew Y. Ng, Priya L. Donti, Lynn H. Kaack, Kelly Kochanski, Alexandre Lacoste,  Kris Sankaran, Andrew Slavin Ross, Nikola Milojevic-Dupont, Natasha Jaques, Anna Waldman-Brown, Yoshua Bengio, Alexandra Luccioni,  Tegan Maharaj, Evan D. Sherwin, S. Karthik Mukkavilli, Konrad P. Kording, Carla Gomes, Demis Hassabis, John C. Platt, Felix Creutzig and Jennifer Chayes.

Continue to Part 6: Datasets & further resources (available from Friday, Sept. 27th)

 

Paul Strobel

Paul is passionate about computer vision and neural networks for image recognition. In addition, he likes to create explanatory videos to easily convey complex content through stories. In his spare time he likes to play volleyball and guitar and occasionally does light painting photography.

Comment

Your email address will not be published. Required fields are marked *