Machine learning could yield faster, cheaper lithium-ion battery development

Scientists have developed a machine learning method that could dramatically slash the cost and energy required to develop new lithium-ion batteries that the modern world is becoming increasingly reliant. Predicting a new battery design's lifespan – and its engineering applications – is a major industry bottleneck. Brute-force …

[_wojtek]

This is actually sensible and useful

This make more sense instead of spitting shitload of crqp by force-fed AI all over the place.

While it's probably not perfect, allowing decreasing research and trial times is very welcomed.

[IGotOut]

Re: This is actually sensible and useful

That's why they were careful to use ML rather than AI. They don't want to be associated with slop.

[hoola]

Re: This is actually sensible and useful

Yes in principle, we need to move away from lithium particularlyfor tractionand storage. . Salt batteries are now viable in terms of capacity and size, they just need the volume to reduce cost.

For small portable devices lithium is still likely to be the better option for now.

Lithium is a complete environmental disaster from source thorough to disposal and recycling.

[Ken Shabby]

Re: This is actually sensible and useful

So if I had a salt battery in my mobile, it would be a Saxaphone? Cool.

[Anonymous Coward]

Re: This is actually sensible and useful

"While it's probably not perfect, allowing decreasing research and trial times is very welcomed."

So you want to be the tester instead?

"not perfect" or "AI tested" isn't an excuse when your battery explodes.

It's all about profits and anyone who believes otherwise, is a bit naive. Sorry.

[jake]

It could, could it?

How about getting back to us after all those coulds turn into something that's functional?

_{At the moment, all this is is a plea for more research grant money to keep researching vapo(u)rware ...}

[Zack Mollusc]

Yeah, no.

I think they are forgetting why we test things.

You think it will be fine, then you test to make sure.

If you think it should be fine and an AI or whatever agrees that it should be fine, you are still at the 'should be' stage and still need to test.

[Neil Barnes]

Re: Yeah, no.

Yabbut, the way to test batteries is in the wild. You do tests for basic operation, safety, and capacity, and then you sell them to people cheaply. With the provisio that they may be dead in a month, or five years, and keep an eye on how they're charging, storing, and discharging.

You might want to consider some sort of secure and fireproof housing for them, of course.

[nobody who matters]

Re: Yeah, no.

".....and keep an eye on how they're charging, storing, and discharging"

...and spontaneously combusting?

I think they really need to be properly tested in all these ways <before> letting them anywhere near the public at large.

[Anonymous Coward]

Re: Yeah, no.

If the "AI" suggests novel options and identify potential failures/successes before testing, then you can focus on just testing likely options that will advance things.

Filtering out the false positives/negatives is what the testing will do but because you have reduced the number of things you test and with "higher" potential of success then you can afford to be more inventive with what you actually test and the rewards could be greater. We are a long way from trusting AI results, but even an "untrusted" system can help if it speeds things up with better results than the traditional approaches.

Of course the elephant in the room is can you develop models that will give you this in the domain - from one presentation from a company working with (selling) us AI (some useful, some vapour), there are some models in battery technology that are progressing things (though the cynic in me says never trust a salesman).

[steelpillow]

Saving over 90% of "half the annual electricity generated in California" equivalent? Are we really spending that much on it?

I think these guys are just after a new Reg standard unit.

[retiredFool]

I'm not sure how they are testing. With a number like 130TWh, I'd hope they have a grid connection and the testing involves actual grid usage. Tightly monitored, in an explosion proof area. Sounds like they just have it hooked to a giant resistor for the discharge. And I mean giant.

[DS999]

The amount of energy consumed in one research lab in a university might only be a few hundred kW, but if that is being consumed essentially 24x7 that's multiple GWh per year. Then you have how many universities and corporate labs all over the doing this, plus it is easy to imagine there are some who are testing on much larger scales than mere hundreds of kW.

If you we can abandon some tests early based on machine learning that says "this pattern of early behavior has always led to poor long term results" then you either save energy by dropping long term testing on 90% of candidates instead of seeing them through to the end, or you're able to test more combinations at the same consumption thanks to early abandonment of pointless tests and overall progress happens more quickly.

[elsergiovolador]

AI

Hey AI can you make us happy, feeling loved and rich?

Short answer: no. I am code, not a fairy godparent with a trust fund.

There you go folks. I asked.

[Anonymous Coward]

'Could' !!!???

The operative word is 'could'.

Although it is ML and not 'AI' it still is a 'could'.

The idea still needs to be tested the 'old' way to confirm the 'could' is valid.

Everyone is looking for 'shortcuts' BUT the world generally does not work that way.

:)

[Jou (Mxyzptlk)]

How to spot scientist

No using buzzwords. They are using the term "Machine learning". Even mention the three parts correctly: Data gatherer, interpreter, predictor, though calling the latter "Oracle module" is a marketing thing :D, but hey, we got a whole Linux distribution chain named after toy story.

And in the end: Promised is a faster statistics result, not the ultimate answer to life, the universe, and everything.

I love scientific types.

Whether it is worth it, we could see.

[elsergiovolador]

Re: How to spot scientist

Is machine learning or having its memory reorganised to provide desired output?

[Jou (Mxyzptlk)]

Re: How to spot scientist

Of course one can always cheat to get the desired output :D.

[Anonymous Coward]

Re: How to spot scientist

Yeah, oracle, the concept is part of both quantum computing and decision problems these days, with the example of Markov Decision Processes (MDP) used in reinforcement learning among others (eg. Q-learning -- that is not neural-net-oriented iiuc).

I understand none of it of course, but hope that whatever the UMich team is doing works like Sandia to help identify some heretofore not-considered aspect of battery design that can foster both capacity and safety while reducing the parameter space over which to perform testing ...

With due care, the generation of Ford Pinto EVs should right be prevented, imho! ;)

[Anonymous Coward]

"cut years from testing cycles"

What they actually mean they'll skip testing claiming "AI says it's OK" and in pratice it's the customers who are the testers.

Pure profit of course. Sole reason to do that, as usual.