Excerpt:
A secretive startup backed by Bill Gates has achieved a solar breakthrough aimed at saving the planet.
Heliogen, a clean energy company that emerged from stealth mode on Tuesday, said it has discovered a way to use artificial intelligence and a field of mirrors to reflect so much sunlight that it generates extreme heat above 1,000 degrees Celsius.
=====
The article describes a bit about the role the AI is playing, and Heliogen didn't seem to me to have used "AI" as a buzzword
Later in the article computer vision and edge detection is mentioned... But I also feel this is more like a traditional control problem rather than ai.
Yeah, the use of "AI" terminology is dubious here. They've replaced aiming via a pre-loaded ephemeris with a system that actually aims the reflected light by looking at where the light is ending up after being reflected and adjusting accordingly. There may be some machine learning going on in the sensing, but this feels an awful lot like basic feedback loop automation.
My biggest pet peeve is when AI is used to describe optimization models, and then people complain of black box algorithms, when the optimization is, in fact, perfectly transparent.
"Heliogen uses computer vision software, automatic edge detection and other sophisticated technology to train a field of mirrors to reflect solar beams to one single spot."
In principle I think you are right, but that would mean for each plant you build you need the exact position of each mirror, the oven chamber and the motion of the sun through the day and year. I think the "AI" approach scales better, as it is rather simple "AI" application.
wouldn't just giving all of the mirrors survey quality GPS do the same. you'd be able to then just have them work out the angles as it relates to them and the target.
It doesn't really seems like and AI problem.
I had no idea what RTK was before your comment. Thanks! Here's a link[1] for others who might be curious. Looks like the key is being within ~6mi of a RTCM (Radio Technical Commission for Maritime) station that transmits GPS correction data. If not, the cool thing is that you can create your own station as well by daisy-chaining off of an existing station that's within 6mi.
Ssh, hacker news hates simple solutions like fiducial markers. I wonder if HN would dare to ride a spaceship that uses neural network to perform docking alignment.
From my reading of the article, one of the big issues to overcome was cost of CSP vs Photo Voltaic . They're even using smaller flat mirrors that are less sophisticated since per panel costs add up fast. One of the advantages is that this takes 4 cameras across the full plant, not a per panel cost.
That's just my reading of it, not asserting anything myself.
>or each plant you build you need the exact position of each mirror, the oven chamber and the motion of the sun through the day and year
Yeah, depending on the level of precision required, there are _tons_ of variables that could have an impact here. Year to year the angle of the sun changes slightly on any given solar date, as various components of the Earth's orbit "wobble" (axial tilt, eccentricity, precession, etc). Even on an hour by hour basis, thermal expansion could have a significant impact. Any significant structure is going to settle to some extent over time, etc...
The motion of the Earth in space is very well determined and studied; the thermal expansion and deformation of the structure is most likely the dominant source of uncertainty for the pointing.
I am wondering if these effects are really important and warrant the use of "AI". If so, then what is the size of the target? If it is only some cm^2, then they still have a long way to go till it can handle industrial sized problems. Once you are at industrial scale, do you still need that kind of accuracy??
I don't think it's about warranting "complex AI" it's that this solution plays out as easier to apply in practice.
I think it's a scale issue, each mirror needs the same apparatus and AI, but doing it traditionally with reactive optics would need a multitude of different pieces?
Ground subsidence and related structural effects on the mirrors and mirror field are the main issue.
But also the position of the sun in the sky is not predictable enough to do this completely open loop because of variances in atmospheric refraction... but it's pretty easily measured at the site.
I'm just going by the HelioMax blurb, where it talks about the calibration time. It makes sense that a system that automatically achieved accurate mirror aiming would also be able to maintain higher temperatures across varying conditions (and recover from disturbances and so on).
Perhaps they adjust the mirrors to optimize the power input in a similar way as gradient descent, and this is why they call it "AI", or something along these lines.
To solve the geometry problem, you need to know the inputs, i.e. position of the sun, mirror, and the target. Obvious way is calibrating in advance but that’s not reliable enough, due to wind and thermal expansion acting on the target (it’s elevated) and seismic activity acting on the ground. Another thing, need accurate orientation sensors in the mirror and/or motors capable of precise absolute positioning. All these technical challenges are solvable, but they reduce reliability, reduce focus precision especially over time, add non-trivial costs to the installation.
Instead, apparently they have 4 fixed mounted cameras around the target, looking downwards at the center of the mirror array. The sky doesn’t have distinct features unless clouds, but the sky has these color gradients, vertical one from horizon to zenith, also radial one around the sun. Just seeing these 4 colors might be sufficient to correctly orient the mirror.
That's a research/materials testing facility. Heliogen is aiming for industrial applications which means it has to scale more cheaply than a bespoke installation.
The sun might not, but everything else does. The struts supporting these things change with the temperature, the ground swells or settles, the tower may move a bit, etc.
Also, once the AI is written, you can save time on each install by not having to precisely locate each mirror. Get it close enough, and let the magic box fix up the small errors in real time.
> The problem is that in the past concentrated solar couldn't get temperatures hot enough to make cement and steel.
> While other concentrated solar companies attacked this temperature problem by adding steel to make the technology stiffer and sturdier, Heliogen and its team of scientists and engineers turned to artificial intelligence.
> Heliogen said it is generating so much heat that its technology could eventually be used to create clean hydrogen at scale.
One of my discarded project ideas was exactly this, but on a smaller scale. A steerable mirror on a mount with a camera for targeting. The idea was to place one on the ground outside a north facing window. The device would then busy itself reflecting the sun up into the room making a nice sunny patch on the ceiling to light the room. They could also be used to deliver extra light to windows with plants in them to grow plants which require more light than is available or to modulate the daily light to fake a different season.
Sitting in 2020 the technology is near trivial. The problem is that some knucklehead would point a bunch of them at the same window and set curtains and pigeons on fire. (You’d do this because you had a large room with few windows, so dispersed mirrors would give you multiple dispersed lighting spots in the room… and incinerated pigeons.)
I was unable to make the leap to what happens if you are a deliberate and well reasoned knucklehead! In a controlled situation a megawatt of light might be just the thing.
Addendum: The mass field of mirrors introduces a lot more work. Some of the immediately interesting problems:
• you need a “everyone stop shining at the target” mode for emergency stop, but the mirrors need to move away from the target in such a way that they don’t create convergences elsewhere. Obviously sweeping your portal to hell down the tower would be a bad idea. So you have a collection of autonomous, free thinking mirrors, but when commanded to stop hitting the target they need to not make the same decisions.
• your camera probably needs to be able to stare into the heart of a thousand suns (literally). But it also needs to see in ordinary daylight. Maybe a pinhole filter that moves over the lens would be appropriate.
• with 999 mirrors shining on the target, you will not be able to see your own contribution to aim. There probably needs to be a “hey! Everyone else look away for a second” command for unit to aim and calibrate itself. See complexities from the first bullet and not incinerating things you like.
So yeah, they could have started with my garden heliostats, but then it would be a crack the knuckles, bring in the eggheads, and get to work operation to solve all the hard problems.
Ok, considering someone had such good idea which is rather obvious in hindsight, but no one had it yet, what would you do to get at least some good money from that, when you don't have enough money to invest in it yourself?
Write an outrage-porn blog post about it and collect ad revenue. But seriously people come up with ideas simultaneously all the time but usually only one of them is getting credit
How to reward inventors who do not have the means or desire to commercialize their work is precisely the problem addressed by the licensing of patents. In practice, especially in computer related fields, the system isn’t working out so well. That’s a problem for governments to solve for humanity, or not and live in a world where independent inventors can not command compensation.
Add similar software to a swarm of drones each carrying mylar film stretched on a lightweight rigid frame with a vertical tilt actuator, and you have a seriously dangerous weapon system on sunny days.
There's a Norwegian town that did something similar: During winter the mirror (mounted on a hill) reflects the sun to the town's market square, because otherwise there wouldn't be any sunlight due to the sun being too low. https://www.theguardian.com/world/2013/nov/06/rjukan-sun-nor...
A sun angle sensor is great for pointing at the sun. Heliostats need to bisect a ray to the sun and one to the target. Installation requires somehow setting the angle to the target. Then you immediately become aware of how much the ground heaves throughout the day, so you have to make a foundation that goes well down into the earth. Now you have a $200 heliostat on a $1000 foundation with a $100 calibration operation.
Using $50 of computer and camera to steer instead of photodiodes and analog controller (presumably in a splayed viewport configuration) lets you lose the $1000 foundation and simplify the calibration. It’s no longer a precision operation, just a pointing out the target.
No, with the photodiodes you can still steer the mirrors. The mirror part doesn't change.
Nothing about the installation changes, except you throw out the "AI" and "computer vision" hype.
How do you know where the target is? That's the part where a computer helps. You don't need to see to know where the sun is given a frame of reference, that's just math and a clock. Your frame of reference is initially unknown and varies slightly, but significantly, with time. Training a computer vision system to look for "insanely bright rectangle subtending <1° of space elevated on a stick" is just the thing computer vision is becoming good at. It's even easier than asking my phone to "show me my pictures of goats" (111 hits, zero false positives (assuming sheep are goats)). Of course, a tall skinny field tech should think twice about wearing a short, bulky, white jacket.
I somehow doubt Gates gets fooled by hype. Do you think he just stumbled his way to success? Maybe there is a little more complexity to the problem than you anticipate.
I tried to ask this the last time Heliogen was on HN, but still don't understand the solution, so I'll ask again:
You have four photodiodes surrounding the target, and 1000 mirrors aligned to reflect the sun toward that target. One of the 1000 mirrors goes out of alignment, and your photodiodes tell you so. How do you know which mirror needs to be adjusted?
A reply into oblivion… but generally you would circulate molten sodium through the system to transfer heat to ground level where you do work and have storage tanks for hot sodium when the sun isn’t shining.
How much energy is required to build mentioned sodium system, how long does the sodium system last? The tree is the best solar panel, and nature invented it.
Being downvoted for an unsubstantiated blanket claim like that doesn't even begin to suggest a bias. The only way such a comment wouldn't be downvoted would be if there were a widespread bias against the person in question.
Pick a random person. If you posted this exact statement about them, would you expect to be downvoted? Personally, I'd downvote it regardless of who it was.
I think the article is overplaying the promise of CO reductions.
Around 60% of CO2 (see 1) emissions during cement manufacturing are just a byproduct of the chemical process converting calcium carbonate to lime (see 2) - nothing that solar power can help with.
I thought that the problem with air travel is not (just) the amount of CO2 emitted, but also the altitude, which makes the green house effect bigger per ton of CO2.
It's not about CO2 emissions, but rather the high-altitude NOx that contributes more to climate change than low-altitude emissions. However, it's debatable whether the amount is very significant, at least compared to other sources of NOx.
The byproduct of the chemical process is largely reabsorbed as the cement cures. The part that's from the fuel burned to heat it is not, and that's the part that matters in the long-term balance.
Another interesting problem this might solve is birds.
Yes, domestic cats kill orders of magnitudes more birds, but that argument a. depends on relative scale that people don't grok, and b. coming from an energy company reinforces the preconception that they're callous.
Add radar and a sufficiently smart AI, and they could turn a PR problem into a marketing pitch by showing birds flying through unscathed while the mirrors shift to avoid them.
If you can track individual birds like that I'd have thought you could do targeted scaring to deter the birds before they get fried. Moving mirrors quickly enough seems like it would require quite significant adaptation of the apparatus?
I worked with a security company that did essentially that with air canons and radar. Unfortunately, scared birds aren't the most predictable creatures, but it seems to work pretty well.
Backing up solar production with batteries is pointless. Except for a really specific cases. The costs are too high, the stress on the grid is too high, the grid need to be overscaled, the batteries need to be changed each 15 years the solar panels need to be changed each 20 years.
And even if we overcome this, we will need so much solar panel and batteries that we simply don't have enough raw materials to do so (to replace a 1GW nuclear core you need to cover half of Paris with solar panels, source http://www.assemblee-nationale.fr/15/rap-enq/r2195-t1.asp).
In France, 95% of the stored electricity is done using reversible-damns (you pump water up to store, you empty the water and generate electricity when you need it), only 2% using batteries. We basically put damns everywhere we could on the territory. We have 2 days max of storage. How can we now store weeks/months of solar-panel/wind electricity we don't need to use it when we need it?
Batteries are inefficient for grid storage, oversized, heavy and too short term. That's it. We now know for decades that damns are the easiest way to do it and we can't do more than a few days of storage.
At one moment we simply have to face the maths and physics. Trusting the "next big thing" in battery storage/solar production, even if we do 5x more energy, even if we develop long lasting ones, will not work in the end. Our needs are too big and the amount of minerals and other raw materials we need is way too big. That's it.
Yeah true. I heard of another Bill Gates backed startup that actually stored the energy in a form of potential energy which was way more cost effective. Keep in mind though dams are super catastrophic to the environment as far as water flow which can be detrimental to moisture and animals that depend on water in certain areas
I'm sure you could do it with a closed system, e.g. pump water up a cliff, and when you need power, bring it back down. I'm guessing using dams on existing rivers is just more convenient than building a closed system from scratch since it can double dip on hydro infrastructure.
Am I missing the breakthrough? This just reads like boilerplate tech marketing, replete with questionable ai usage. I suspect without its well known backer anyone would notice.
They seem to be using computer vision techniques to dynamically aim the mirrors. Presumably static or open-loop techniques don't focus as precisely, and using a closed loop also allows them to reduce the stiffness (and this material use) in their mounts and/or fewer mirrors.
Pointing hundreds of mirrors is expensive. Requiring a lot less rigidity for the motion system is notable.
I don't doubt you could have done this open loop with a very rigid (expensive) platform before (including solving problems of ground subsidence). But now you can do it with a reasonably priced platform closed loop, which is what makes it actually possible.
Open loop means that they’re just using a mathematical prediction of where to point the mirror at a specific time, given the relative position between the heliostat, the target and the sun. adjustments to the mirrors are typically done every 10 to 15 seconds, which is sub optimal. Closed loop means that they’re actually using Computer vision to dynamically sense the specific angles between the sun and the target and then adjust the mirrors accordingly, no mathematical model required, and the adjustments can be done tens of times per second.
Yah, it's a slight misnomer because both loops are likely closed.
Open/closed loop basically means whether you check the result of an actuation. What we're really talking about is the point at which the loop is closed.
A completely open loop would run the motor for the predicted time.
A simple closed loop would run the motor until the mirror is pointing at the predicted angle, using an encoder to know where the mirror is pointing. The control loop encompasses a mechanical part of the system, but not the end-thing we want controlled. Or put another way, the loop is closed for actuator angle but is not closed for where reflected light goes.
A better closed loop would run the motor until the mirror is reflecting sunlight exactly where we want it-- using magic to know where the light from this mirror is going.
Makes sense. So closed loop basically means there's a feedback mechanism in place to ensure that the input results in the desired output, or some sufficient proxy thereof.
We have beed dynamically controlling mirrors in closed loop since mid seventies - year of the first CDROM drive. You dont need trendy technobabble like AI for this task.
Yes you are. High temperature from concentrated solar power has been something of a holy grail for a while. How it was achieved ("with AI") is less interesting than the result.
If you watch the interview, the founder of the company never uses the word AI. He uses the words "software" and "computer controlled". However, he does say "yes" when the interviewer asks "So you use AI?".
106 comments
[ 2.8 ms ] story [ 166 ms ] threadHeliogen, a clean energy company that emerged from stealth mode on Tuesday, said it has discovered a way to use artificial intelligence and a field of mirrors to reflect so much sunlight that it generates extreme heat above 1,000 degrees Celsius.
=====
The article describes a bit about the role the AI is playing, and Heliogen didn't seem to me to have used "AI" as a buzzword
In principle I think you are right, but that would mean for each plant you build you need the exact position of each mirror, the oven chamber and the motion of the sun through the day and year. I think the "AI" approach scales better, as it is rather simple "AI" application.
If every mirror has drifted a small angle, and you're trying to point them all at a very small target-- it's hard.
[1] https://learn.sparkfun.com/tutorials/what-is-gps-rtk/all
That is, with two GPSes, you can measure their vector distance very precisely by correlating their measurements.
But it doesn't help you precisely measure angles, which is a big part of this problem.
That's just my reading of it, not asserting anything myself.
Yeah, depending on the level of precision required, there are _tons_ of variables that could have an impact here. Year to year the angle of the sun changes slightly on any given solar date, as various components of the Earth's orbit "wobble" (axial tilt, eccentricity, precession, etc). Even on an hour by hour basis, thermal expansion could have a significant impact. Any significant structure is going to settle to some extent over time, etc...
I think it's a scale issue, each mirror needs the same apparatus and AI, but doing it traditionally with reactive optics would need a multitude of different pieces?
But also the position of the sun in the sky is not predictable enough to do this completely open loop because of variances in atmospheric refraction... but it's pretty easily measured at the site.
https://heliogen.com/
I'm just going by the HelioMax blurb, where it talks about the calibration time. It makes sense that a system that automatically achieved accurate mirror aiming would also be able to maintain higher temperatures across varying conditions (and recover from disturbances and so on).
To solve the geometry problem, you need to know the inputs, i.e. position of the sun, mirror, and the target. Obvious way is calibrating in advance but that’s not reliable enough, due to wind and thermal expansion acting on the target (it’s elevated) and seismic activity acting on the ground. Another thing, need accurate orientation sensors in the mirror and/or motors capable of precise absolute positioning. All these technical challenges are solvable, but they reduce reliability, reduce focus precision especially over time, add non-trivial costs to the installation.
Instead, apparently they have 4 fixed mounted cameras around the target, looking downwards at the center of the mirror array. The sky doesn’t have distinct features unless clouds, but the sky has these color gradients, vertical one from horizon to zenith, also radial one around the sun. Just seeing these 4 colors might be sufficient to correctly orient the mirror.
Animats gave a nice summary [1] of the Vox article [2] that went into some detail. [1] https://news.ycombinator.com/item?id=21580596 [2] https://www.vox.com/science-and-health/2019/11/19/20970252/c...
Could probably find enough information here [0] what is needed.
[0] Practical Astronomy with your Calculator https://www.amazon.com/dp/0521356997/ref=cm_sw_r_cp_api_i_-v...
Also, once the AI is written, you can save time on each install by not having to precisely locate each mirror. Get it close enough, and let the magic box fix up the small errors in real time.
> The problem is that in the past concentrated solar couldn't get temperatures hot enough to make cement and steel.
> While other concentrated solar companies attacked this temperature problem by adding steel to make the technology stiffer and sturdier, Heliogen and its team of scientists and engineers turned to artificial intelligence.
> Heliogen said it is generating so much heat that its technology could eventually be used to create clean hydrogen at scale.
Sitting in 2020 the technology is near trivial. The problem is that some knucklehead would point a bunch of them at the same window and set curtains and pigeons on fire. (You’d do this because you had a large room with few windows, so dispersed mirrors would give you multiple dispersed lighting spots in the room… and incinerated pigeons.)
I was unable to make the leap to what happens if you are a deliberate and well reasoned knucklehead! In a controlled situation a megawatt of light might be just the thing.
Addendum: The mass field of mirrors introduces a lot more work. Some of the immediately interesting problems:
• you need a “everyone stop shining at the target” mode for emergency stop, but the mirrors need to move away from the target in such a way that they don’t create convergences elsewhere. Obviously sweeping your portal to hell down the tower would be a bad idea. So you have a collection of autonomous, free thinking mirrors, but when commanded to stop hitting the target they need to not make the same decisions.
• your camera probably needs to be able to stare into the heart of a thousand suns (literally). But it also needs to see in ordinary daylight. Maybe a pinhole filter that moves over the lens would be appropriate.
• with 999 mirrors shining on the target, you will not be able to see your own contribution to aim. There probably needs to be a “hey! Everyone else look away for a second” command for unit to aim and calibrate itself. See complexities from the first bullet and not incinerating things you like.
So yeah, they could have started with my garden heliostats, but then it would be a crack the knuckles, bring in the eggheads, and get to work operation to solve all the hard problems.
I don't really know if this would be possible, but sailing ships can rig their sails to propel them against the wind.
Of course it worked the first time they turned it on - they were using 4 photodiodes. The "AI" and "computer vision" is just to get funding and hype.
I'm going to start taking solved problems and applying BS "AI" to them to get funding, sheesh.
Using $50 of computer and camera to steer instead of photodiodes and analog controller (presumably in a splayed viewport configuration) lets you lose the $1000 foundation and simplify the calibration. It’s no longer a precision operation, just a pointing out the target.
"Bill Gates is smart, therefore he can't be fooled or wrong" is not a strong argument.
But yes, fiducial marks probably don't work well here due to too many variables.
You have four photodiodes surrounding the target, and 1000 mirrors aligned to reflect the sun toward that target. One of the 1000 mirrors goes out of alignment, and your photodiodes tell you so. How do you know which mirror needs to be adjusted?
Everything Bill Gates has ever done (or financed) has been stolen, a fraud or both.
Edit: Whoa, heavy bias towards Mr Gates, a surprise dear HN community. I'm using my karma to learn! ;)
In this case he is just the investor.
Pick a random person. If you posted this exact statement about them, would you expect to be downvoted? Personally, I'd downvote it regardless of who it was.
Around 60% of CO2 (see 1) emissions during cement manufacturing are just a byproduct of the chemical process converting calcium carbonate to lime (see 2) - nothing that solar power can help with.
1) https://precast.org/2013/05/concrete-and-co2/ 2) https://www.ipcc-nggip.iges.or.jp/public/gp/bgp/3_1_Cement_P...
[1] https://en.wikipedia.org/wiki/Environmental_impact_of_aviati...
Yes, domestic cats kill orders of magnitudes more birds, but that argument a. depends on relative scale that people don't grok, and b. coming from an energy company reinforces the preconception that they're callous.
Add radar and a sufficiently smart AI, and they could turn a PR problem into a marketing pitch by showing birds flying through unscathed while the mirrors shift to avoid them.
See: https://newatlas.com/optical-illusion-raptors-airplanes/5675...
And even if we overcome this, we will need so much solar panel and batteries that we simply don't have enough raw materials to do so (to replace a 1GW nuclear core you need to cover half of Paris with solar panels, source http://www.assemblee-nationale.fr/15/rap-enq/r2195-t1.asp).
In France, 95% of the stored electricity is done using reversible-damns (you pump water up to store, you empty the water and generate electricity when you need it), only 2% using batteries. We basically put damns everywhere we could on the territory. We have 2 days max of storage. How can we now store weeks/months of solar-panel/wind electricity we don't need to use it when we need it?
Batteries are inefficient for grid storage, oversized, heavy and too short term. That's it. We now know for decades that damns are the easiest way to do it and we can't do more than a few days of storage.
At one moment we simply have to face the maths and physics. Trusting the "next big thing" in battery storage/solar production, even if we do 5x more energy, even if we develop long lasting ones, will not work in the end. Our needs are too big and the amount of minerals and other raw materials we need is way too big. That's it.
For electricity production, photovoltaics is superior to concentrated solar power.
I don't doubt you could have done this open loop with a very rigid (expensive) platform before (including solving problems of ground subsidence). But now you can do it with a reasonably priced platform closed loop, which is what makes it actually possible.
Open/closed loop basically means whether you check the result of an actuation. What we're really talking about is the point at which the loop is closed.
A completely open loop would run the motor for the predicted time.
A simple closed loop would run the motor until the mirror is pointing at the predicted angle, using an encoder to know where the mirror is pointing. The control loop encompasses a mechanical part of the system, but not the end-thing we want controlled. Or put another way, the loop is closed for actuator angle but is not closed for where reflected light goes.
A better closed loop would run the motor until the mirror is reflecting sunlight exactly where we want it-- using magic to know where the light from this mirror is going.
In practice, you often end up with multiple nested control loops on this kind of system, each of which is closed.
You might have a voltage controller deciding how much to turn on the motor transistors based on voltage feedback...
You might have a motor speed controller which decides what voltage to give the motor based on speed feedback...
With another position controller deciding what speed the motor should try to go based on position feedback...
And then the magical-mirror-aiming thing deciding where the position controller should try to point it based on other feedback.
And then the industrial process controller deciding how many mirrors should be pointing at the thing based on temperature feedback...
And some other controller deciding what the temperature should be based on process feedback...
And then humans turning the knobs on that controller based on QA feedback.
(And each of these control loops usually gets at least 2x slower as the one it encloses).
Kinda like back when everyone started moving to the web...