Tackling Climate Change with Machine Learning: the Due Diligence

ML can be an invaluable tool both in reducing greenhouse gas emissions and in helping society adapt to the effects of climate change. Climate change is a complex problem, for which action takes many forms – from designing smart electrical grids to tracking deforestation in satellite imagery. Many of these actions represent high-impact opportunities for real-world change, as well as being interesting problems for ML research.”

A community of researchers including OpenAI, DeepMind, Element AI together has started to explore the role that “machine learning can play in mitigating and adapting to climate change. They have published a research agenda, started to collaborate with climate experts, inspired faculty to develop courses on climate and ML, and organized workshops at the major ML conferences such as ICML and NeurIPS.”

Here from limited my purview, with the help of friends and a climate scientist, here are my thoughts on the four research ideas proposed to combat the greatest climate crisis of our generation.

Initial Thoughts:

  • I thought those were all very good ideas grounded in solid research with a team of unparalleled caliber.
  • $2 million may not be enough. A startup can burn $2million quickly, even in one year using Amazon Web Services. This would require a very good engineer to build a beautiful system, that would mean competing for someone with well-rounded startup experience OR 10 engineers in SF with salaries ranging around 150K to $2million/year easily.
  • Calling these tech/innovative components may not create the urgency for the burn rate for each the startups, which I think would be $500K each, which is doable, but needs entrepreneurs that can grind with the money. Great salespeople.

Here are the tech ideas:

  1. Improving energy efficiency of buildings by using ML to interpret building occupancy data, and reducing energy consumption on HVAC and lighting; 

The biggest problem in HVAC is energy waste is from dissipation (poor insulation or non-passive designs) and lack of renewable alternative fuels (heating). Renewable energy sources are generally very sporadic in power generation say 30% uptime, but the main solution so far has been to have fundamental breakthroughs in battery storage that could efficiently store and use locally the DC (generated via the solar/wind plants – local usage think powering the lights in remote mines that is right next to and connected to the DC power source), convert into AC and power other parts of the surroundings or even feedback into the grid. It’s a storage and/or distribution problem on a physical/chemical level and not too much of an info asymmetry level that feeding data overlaying utility rates can be solved.   

If forecasting is the best landscape for say wind/hydro generation entirely from a power generation efficiency point of view this application of ML can be interesting in improving planning because for hydro/wind. This can take years to construct and plenty of “collateral” geographical impact so proper prior planning prevents poor performance due to the high CapEx cost. Right application of data learning to process immense datasets and speed up/improve decision making in an expensive “wrong decision” environment.

2. Using model-based reinforcement learning for perishable inventory management to reduce the GHG emissions associated with food waste by supermarkets by 30 percent;

Supermarkets do not have a huge margin, so they may not be motivated to adopt new tech. There is an app called Damago, run by a Korean American doing this – a food app that connects customers with unsold menu items – at a discounted price.

3. Improving forecasting of solar and wind power forecasting, which increases the capacity of utilities to use solar and wind without undermining the reliability of the grid;

 4. Supporting the Measurement, Reporting and Verification needed to pay developing countries for forest conservation, using “interpretable ML” to analyze satellite images.

Ah yes, but also costly. Most of AI/ML is geotagging/grunt work – requiring CNN (convolution neural network) tech knowledge. There is a great startup by a close friend named Karina, former UNDP Consultant. She runs TaQadam with former Syrian refugees essentially geotagging satellite imagery data making them “AI-ready” in other words, much cheaper. But there are a lot of attempts to do this to improve satellight imagery data, I think also by the CTBTO. 

Forest conservation a really tough nut to crack because it involves the basic livelihoods for almost all commodity/land driven economies + political stability of the nation as a whole involving the employment of locals. A wide variety of economic-political factors are in play here so not something that a market enterprise approach like paying a dollar off to Indonesia’s lost land-related GNI can solve. 

Even if images were ready & AI definitively incriminates Indonesia/Brazil of killing Earth’s lungs, they’re not going to shut down their palm oil plantations or cattle ranches. And no developed country is going to be able to sustainably pay them to stop. for this one, AI will only confirm what we already know, a sustainable development problem. Real choke point is smoothing supply/demand mismatch cheaply, a la cheaper battery. battery chemistry problem.