If you are in a rush, read this first
I used Mergeflow’s analytics to search for venture investments, R&D, and patents at the intersection of “climate change solutions” and “machine learning technologies”.
My starting points were a recent study by some leading machine learning researchers, and an article in MIT Technology Review by Karen Hao, on how machine learning could be used to tackle climate change.
I found that…
…”computer vision” is used by quite a number of venture-capital-funded companies, across climate change solutions.
…”machine learning for carbon capture” is an area with R&D but almost no patenting activity (yet).
…of the top-10 highest-growth R&D topics across “ML tackling climate change”, 8 are directly relevant to electricity systems.
If you have a bit more time on your hands, continue here
In June, a group of machine learning researchers, led by David Rolnick, published a study on tackling climate change with machine learning. From this study, Karen Hao distilled ten recommendations where machine learning may have a particularly great impact. Karen’s article in MIT Technology Review included examples such as “improve predictions of how much electricity we need”; “discover new materials”; “optimize freight routing”.
Besides specific recommendations, the full study as well as Karen Hao’s article also feature a matrix that maps climate change solution domains against various machine learning technologies:
“This matrix looks like the data we generate with Mergeflow’s Grid Search!”, I thought.
“What is Grid Search?”, you ask.
Mergeflow Grid Search is a component of our analytics platform. It lets you systematically explore large numbers of topics. So, rather than you having to manually search for R&D, investments, patents, etc. for every one of your topics, and then for the combinations of your topics, Mergeflow automates this for you.
For this article, I used Mergeflow’s Grid Search to explore R&D, venture fundings, and patents at the intersections of climate change solutions and machine learning technologies.
I wanted to…
- …discover venture-capital-backed companies that use machine learning to fight climate change.
- …compare R&D and patents: where do we see R&D but no patents (yet)?
- …find out which machine learning / climate change solution combinations have seen the biggest growth in R&D.
How I searched…
First, I built a matrix, “climate change solutions” x “machine learning technologies”, using the outline of Rolnick et al’s paper as a guideline. For example, in climate change solutions, I searched for solutions in electricity systems, transportation, buildings & cities, etc.. In machine learning technologies, I distinguished general topics (e.g. neural networks) from applications (e.g. computer vision) and what I consider “emerging topics” (e.g. one-shot learning or decentralized machine learning). The result was a matrix, mapping climate change solutions against machine learning technologies.
For each row and each column of my matrix, I specified a Mergeflow search query. Mergeflow Grid Search then used my queries to automatically build matrices that show the intersections of climate change solutions and machine learning technologies. I did this for R&D, venture investments, and for patents.
Of course, I could have chosen a different setup for my matrix. I’m not saying that my way is the best or the only way. If you have a different proposal, I’d be very happy to hear from you.
…and what I found
1. Venture-capital-backed companies that use machine learning for climate change solutions
First, I used Mergeflow Grid Search to produce a matrix of venture investments in “climate change solutions” x “machine learning technologies”. You can see the results below (click on the image to see a bigger version). The numbers indicate sums of VC investments in each topic combination.
This picture is a screenshot of an interactive matrix. In Mergeflow, you’d be able to zoom into the matrix, and to click on the numbers to see the underlying data, i.e. the venture-capital-backed companies.
So what does this matrix tell us?
Everybody and their dog (or cat) do ML (or AI).
Or they claim they do. Well, we kind of knew that, right? So let’s not dwell on this.
Computer vision is a cross-sectional machine learning application.
I found this more interesting, so I zoomed in. If you’d like to get the full set of findings, please get in touch. For now, I’d just like to point out two examples of companies that I discovered:
Aurora Solar (https://www.aurorasolar.com/). They use computer vision to automatically determine how many solar panels will fit onto a property, how much energy they will produce, and how much they will bring down the owner’s electricity bill.
eSmart Systems (https://www.esmartsystems.com/). eSmart uses computer-vision-equipped drones for visual asset inspection (assets = power lines, wind turbines, etc.). But they also integrate data from other sensors in order to optimize electric grid resource usage and management.
Interestingly, both companies use computer vision as part of a full-stack solution. “Get solar energy to work for property owners” in the case of Aurora; “optimize my grid resources” in the case of eSmart.
“Reinforcement learning” is making an entry in climate change solutions
While reinforcement learning is not really new, it has received considerable attention more recently. For example, think of DeepMind’s reinforcement learning based AlphaGo beating Lee Sedol, one of the best Go players in the world.
What is reinforcement learning?
Very briefly, in reinforcement learning, the goal is to get an algorithm to master a task. So, for example, rather than saying “here are 500,000 examples of good Go moves and 500,000 examples of bad Go moves” (this would be “supervised learning”), in reinforcement learning, you’d just say, “here are the moves you can do in Go, and whatever you do with those moves, make sure you win the game”. If you’d like to know more, Sebastian Wagner has written a good comparison of “reinforcement learning vs. supervised learning”, for example.
Back to climate change and machine learning. One use case for reinforcement learning that comes to mind here is HVAC (heating, ventilation, air condition). Rather than saying, “here are 500,000 good example HVAC settings”, you could just say, “however you set my HVAC, make sure that carbon emissions stay within my defined limit. And while you’re at it, please also make sure that my electricity bill won’t kill me.”. This is what Carbon Relay, a venture-capital-based company, does. And they do this with reinforcement learning.
2. R&D but no patents (yet)
Next, I wanted to know if there are “climate change solutions” x “machine learning technologies” combinations where we see published R&D but no patents (yet).
This is a bit tricky. I don’t want to not find any patents for a topic combination simply because patents do not use a certain terminology. For example, if patents don’t say “reinforcement learning” but something else instead to refer to the same thing, I might miss patents simply because of a terminology issue. So I made sure I controlled for this issue (basically, by checking first if my search queries match any patents at all).
But why do I even care if there are topics with R&D but no patents yet? Because this could be emerging topics that have not really caught on commercially yet. Different question of course if they ever will. But just in case, I could put such topics on my watch list.
Turns out that “machine learning for carbon capture” might be such a topic. There is one patent, but this patent is for creating carbon capture technology roadmaps, not for actually doing carbon capture. But R&D seems to be increasing a bit. The screenshot below from Mergeflow shows the share of R&D publications over time that relate to “carbon capture and neural networks”:
A lot of this work uses neural networks to try to predict various properties of CO2. For example, the extent to which CO2 can dissolve in various other substances, such as salt solutions (e.g. here).
3. High-growth R&D topics
Moving on. What are high growth R&D topics in climate change and machine learning? To find out, I used Mergeflow’s Delta-t component. This component basically calculated growth rates (CAGRs but quarter to quarter, so CQGRs) for all my topic combinations in my climate-change-machine-learning-matrix.
It’s mostly about neural networks and electricity systems, it seems (time frame is 5 years here):
The CQGR metric basically rewards growth over the whole 5-year time frame that I looked at. In order to reward more recent growth, I used a different metric that simply divided topic volume of the last three quarters by the volume of the previous time period:
Still a lot of electricity systems. But now we see “urban planning” x “neural networks” (marked green) really picking up, for example. Topics there are diverse, ranging from vehicle classification in cities to urban land use analyses and more general urban modeling, for example.
And there is a new entrant here too that has nothing to do with electrical systems: “disease surveillance” and “time series analysis” (also marked green). Here, topics include using Twitter to improve Zika virus surveillance, extracting epidemic symptoms from influenza reports, or analyzing associations between ambient air pollution and risk of type 2 diabetes.
What have I learned?
Looking at my findings, including the ones I wrote about here, I’m happy to see that many of these findings are very concrete. Take Aurora Solar, for example. If I have a solar rooftop installation business, I can subscribe to their service right now. No need to wait for nation states to agree on common climate targets. At times when such high-level agreements are rather unlikely to happen, I think this is good news.