In this interview, Richard Ahlfeld explains his technological approach and how the company has developed in recent years. He explains the algorithms behind his solution and where Industrial AI is heading. We have another job offering: In Heilbronn they are looking for a new professor for “RESOURCENEFFICIENT IT IN HARDWARE AND SOFTWARE”.
Thanks for listening. We welcome suggestions for topics, criticism and a few stars on Apple, Spotify and Co.
We thank our partner **HANNOVER MESSE**
https://www.hannovermesse.de/de/
Job offering: Heilbronn is looking for a professor
https://cdn.hs-heilbronn.de/ce3c0933328f570c/d9e7413e6a15/232-P-IT.pdf
Richard Ahlfeld https://www.linkedin.com/in/richard-ahlfeld/
This podcast is presented by Hanover Messa your leading event for industrial AI hey it’s me Robert as you all know we use the podcast to help our community find new jobs today I have something from hibron and you all know that hibron is one of Germany’s industrial AI centers our listener Professor Dr Robert Alexander vinberg is looking for a new
Colleague a professor for resource efficient it and heart and software you can find the whole job in the show notes and now we will focus on battery testing enjoy Listening hello everybody and welcome to new episode of our industrial AI podcast my name is rber and my guest today is Richard alfeld founder of monolith AI welcome greetings to London welcome and greetings back to you it’s nice to be here Richard please introduce yourself briefly to the listeners what is your
Profession my profession does founder count as a profession I’m not sure I’m not sure you have to explain a bit so I think my background I used to be in Academia for 10 years so I was a in the UK we call it fellow in Germany I think
You might call it an assistant professor at Imperial College London and then I did small research visits in Stanford and at Nasa so I think the best description of my background is a researcher for machine learning and engineering and then my research became a company and so since then I’ve just
Been right a company director an entrepreneur you expression yes so we already had an episode with monolith at the very beginning of our podcast but but today we are talking about batteries because in my opinion you shifted your focus battery is the new hot why I
Wouldn’t say shift it okay i’ rather say we sharpened our Focus because as you might know we built our main product is a machine learning platform that allows any engineering user to build a machine learning model based on sensor data or other engineering product data and people have been using that for all
Kinds of products really for the Last 5 Years and what we’ve noticed gradually I call this the monolith charts batteries have been like moving up in the monolith charts year on year on year and in 2023 they became the number one single that everyone wanted to listen to and they’ve
Become so popular that we said hey look the majority of new deals that we did in 2023 were all on batteries and as you say batteries are the hot right like from battery cell manufacturers popping up left and right in Europe in the US these days to new equipment
Suppliers to new startups being built to new battery test centers we have a couple of solutions here that we’ve built for other purposes over the last years that are very unique that are quite special and they’re very popular and so we’ve said Hey look it’s the right time it’s really motivating it’s
Amazing to work on batteries let’s really push this topic and let’s see how far and how much we can help companies in this area which pain do you do your customers have okay so the pain are like there’s various different forms of pain I think if we start at the biggest
Possible at the high level um then like in the battery in the current situation like there’s a lot of engineering companies who haven’t used batteries for their propulsion system in the past and they now need to switch right and so they first come up against a very simple
Question like which battery should I use and one of the things that is annoying about batteries is that it’s very tricky to know how they will behave you might have experienced this yourself with your smartphone that is very very difficult to tell when the smartphone battery suddenly degrades and I think we
All have this moment after 2 years I could have sworn last yesterday I had a good battery and today I’m out at 2 p.m. like batteries have this very nonlinear sometimes random feeling Behavior where they suddenly start degrading and this behavior is very very difficult to model and understand using physical systems up
Front the only way you can really understand this behavior is by testing it which is why all of the big European Automotive companies are investing hundreds of millions in huge battery test centers and test labs you can imagine this as like a football stadium full of little boxes in which they put a
Lot of batteries and then they cycle these batteries like simulating literally on the hardware prototype on and off and on and off and on and off you the user using this battery on a daily basis and they do this obviously thousands of times to see in what
Conditions at what time does a battery fail so in the past the cars were tested in Finland because of coldness and in the desert and now we have also to test batteries on off on off right correct and it’s funny that you mention sort of Finland in the desert because one of the
Things that makes the battery testing even harder is that like you also have to test them in different temperatures sure and so there is battery cyclist that are just at room temperatures and then there’s battery cyclers so battery things battery tests that go at minus 10us 20 -40 degrees there’s battery tend
To go up to 60° so you’ve got all of those thermal Chambers where you cycle your batter batteries day on day off in different temperatures sometimes for years so this is one pain the behavior right correct understanding the behavior is I think the root cause of what’s
Going to create a lot of other pains because they don’t simply understand the behavior they need to go and test a lot they need to test thousands of batteries in hundreds of different scenarios and now this creates a lot of other problems it creates the problems that it’s very
Hard to decide which battery chemistry I can trust like makes it tricky to pick the right battery because you can’t quite know which specific battery is the best one for your problem so you have to go and test them all and the cycling also creates a lot of data there are
Battery test Labs of a large automotive company might create two terabyte of battery test data in a week and then somebody needs to look at this data and figure out what it actually means and then from this data you obviously also need to derive how the battery behaves
You need to model it you need to understand it and so all of these different things is essentially created fantastic conditions for machine learning M because this is the kind of environment where like there’s a lot of problems in engineering where you look at them and
You think why are you trying to use AI or machine learning it’s pointless you have five data points and nobody knows what’s going on you should just try and figure this out yourself here we have hundreds of different batteries being cycling creating terabyte of data in a
Space that is very hard for any normal person to understand because it’s incredibly nonlinear to make predictions for what millions of people are going to do with their car every day it’s perfect when we first talked at the seens conference I think your marketing slogan
Was test L so to still the approach test less correct it’s still about testing less and I think the reason why we gained momentum in the battery industry is that since 2016 we’ve been developing a toolbox which we I don’t know the internal expression is it’s our Active Learning
Toolbox to explain it to people we usually call it a test plan optimization toolbox okay what this essentially does is that when as a battery cell manufacturer or an automotive EV manufacturer you come up with your plan of like okay this is what we want to
Test here’s our long list of ideas we want to test these temperatures these cycling conditions these driving conditions these voltages and so on so you come with your long list what this test plan feature does is essentially it goes through the list while you start testing and starts figuring out which
Tests do I need and which tests do I not need so it might tell you when you start testing after 3 months you can stop test 27 to 50 because I think you don’t learn anything from those they’re useless you can save your money instead test 17 and
50 I like those are amazing please double down on them you should do more tests in a similar condition because I think we’re learning really important things here and so it’s learning and it’s working actively with a test engineer to figure out what to test to
Learn more and what to test less because they’re not learning anything how you do that let’s take a look under the hood what do you do technically there in your platform okay it now gets tricky because mathematically there’s a variety of different methods that you can do here
So the field of active learning with in machine learning is in principle the entire field of using an existing data set and trying to infer from that what you know about it and what you don’t know about it yeah and there’s a lot of fantastic papers in the battery IND
Industry where for example using basian inference people have been able to come up with what the optimal charging strategy is or what the optimal discharging strategy is or the optimal battery calibration anyway so there is a long list of such algorithms from clustering to random Forest disagreement
To gradient based cing to some forms of reinforcement learning that essentially all are algorithms which look at a set of experiments and try to figure out what you’ve learned from those experiments in order to see what you should potentially do next when I explain it internally to sort of new
Starters at the team I usually tell them do you remember I mean I think maybe five or six years ago there were these type of algorithms that learned how to play Super Mario yes yes and they learned how to play Super Mario basically just by a trial and error like
No real intelligence but essentially just try to play and jump and learned okay if I run into a rock that’s bad I need to go and try it again and they learn by trial and error and this is this class of algorithms which essentially gets rewarded for doing it
Right and it gets punished for getting it wrong and it keeps playing and playing until it gets it right so we use this kind of algorithm class so like self-learning models so it’s a reinforcement approach right in principle yes although reinforcement learning for some of the problems that
Have little data tends to be too inefficient yes in principle it’s I think the best way to understand it is by comparing it to reinforcement learning even though that doesn’t work for all problems but then yes essentially what the algorithm in the background does it’s trying to learn the
Game of how do I test my battery so that my company can understand how to use it in all different conditions as best as possible in as little time as possible and so by giving the engineers who have to test this these algorithms as support it becomes
Essentially from trying to sort of play this game all by yourself which is hard you suddenly try this game and you have a chess computer helping you right and so if you’ve ever tried to sort of win against a chess computer it’s very hard to beat a chess computer a chess unless
You have a chess computer yourself right yes what I find amazing and we’ve seen this more and more throughout this year when we’ve given battery test engine is this machine learning it’s not a bot but like just this recommender algorithm and figure out their test plan that there were first very skeptical because
They’re like well I’ve done design of experiment in University and you usually there’s a de optimal design and you determine the determinant and so on this is how you normally do it I don’t think this algorithm can actually be any better and to be honest I think three or
Four years ago I would have told them they are right but there’s one thing we’ve done it with the algorith GM since then which actually makes it very very useful for test engineers and that is we calibrate it so what we’ve done is that normally I feel like most large
Engineering companies well we’ve tried this sort of active learning and it doesn’t tend to be much better than our engineers and that is true because there’s too many different algorithms and these different algorithms if you put them on a new Problem by themselves they’re not going to solve it if you and
I start now using a reinforcement learning algorithm of the internet or let’s use a random Forest disagreement sampler or anything else and put that on a battery test and compare that against an engineer it’s not going to be better okay because it’s not smart enough it
Hasn’t learned enough so what we do is pre-training so what we’ve essentially done we’ve collected data sets of battery test plans over a while and we’ve built a toolbox that essentially goes through all possible algorithms and starts calibrating itself so so it’s an automl tool a little bit yes so we’ve
Essentially decided we’re not predicting we’re not training a model for your battery test we’re training a model that learns how to best predict battery tests we’re basically training the chess computer before it starts playing so like it’s a pre-trained model that has already learned how to play this game of
Battery testing for a lot of different batteries and so when it comes into the game it’s not the first time it sees this and as a battery engine saying well this is Dum it suggested that I start with room temperature but essentially right it already comes with a suggestion
Of 15 to 20 points where you’re like hm that’s pretty clever it already from the start knew in what conditions this battery is going to degrade faster in a way most experts don’t fully understand so there’s something clever at work here and the next five steps are pretty good
Too so we’ve essentially built a toolbox that a we now have some collaborations where we get a lot of battery data sets so we’ve hoovered everything in yeah that was my question where do your data gathering for the batteries yes so we’ve hoovered everything in we we have a
Collaboration with NASA we’ve just signed one with an Imperial College spin-off similar Imperial College spinoff Ascot about energy and their business model is to go and test batteries for companies and then sell the test data set so they have a library and then are our partners and then we
Have further collaboration shs with other of the test stand providers in Europe one of them will become more public I think in q1 next year so step one gather a lot of the different data sets so we can start pre-training and making sure this works up front and then
The second part is that we still retrain the model when you come with your own problem and this is important okay because battery chemistries change and this is I think an important step because right like just because I’ve learned on a chemistry that was popular three four years ago yes absolutely
Doesn’t necessarily mean I can generate like the same battery is not always going to behave the same way and so when you come with your I don’t know what’s the most fanciful thing that Curr is like a solid state battery like a lithium ion tuned algorithm is not going
To give you the best predictions solid so what we do is that essentially in our software now we have this evaluator function where the person who wants to build a solid state battery can take our test plan Optimizer algorithm upload all of the data that they so far have for
The solid save battery and then the algorithm starts calibrating itself again and trains itself ples the game thousands of times against itself and then comes up with recommendations of what to test so Richard do you sell now a tool box or is it a model what would
You say it’s a toolbox okay because it’s still a lot of things the model itself we’re obviously selling basically both but the toolbox is I think essentially the part that the engineers need because a model in itself is not as useful if you don’t use it for something really
Valuable like validating which tests are useful in figuring out what tests you really need to do to improve the thermal safety of the battery and so there’s right like a lot of different tests and a lot of different applications that you might use this for right like from how
Many charging Cycles what is the best way to cool the battery what’s the capacity what’s the life of the battery there’s so many different things that this algorithm could be doing and it needs to be tuned for all of those different things and so what we’ve done
Is we’ve built this toolbox which allows the engineer to essentially Say Hey I want a learning algorithm for the Aging I want a learning algorithm for the safety and what is unique about this approach and I think that was making so popular is I think everyone in the
Battery industry right now is focusing on building a model of the battery itself all these startups and like their main focus is I will have to build a model of the a big model yeah correct right we on the other hand we’re essentially saying I want to build something that is
Primarily focused on helping you build the model so we’re not competing with people who are saying oh I’m going to build a model of the battery it’s like yes please you should do that and you should probably go and make that a physical model too and not an AI model
Because physical models are interpretable but before you get your physical model you will have to go and test some batteries to validate that it’s actually true that’ll cost you a lot of time and a lot of money and so either way you have to get this testing
Done as fast as possible and then you have to figure out what tests you need to do and this is the only tool that actually helps you figure out what test in what sequence for how long you need to do in order to get there so it’s essentially no matter what approach you
Use physic modeling machine learning model hybrid model it’s accelerating how quickly you get towards having it’s accelerating the time you need to get any model of your battery this toolbox do you deploy it on the edge or on the cloud because you mentioned 100 million EUR in Test
Facilities it’s a area where the automotive companies and the first tier suppliers are very nervous so it’s the edge based design or is it a cloud-based design at the moment it’s a cloud-based design primarily the data sets that come off most of the large battery Labs you’re right they are very very
Confidential highest security and it’s a very confidential area overall but given the size of the data and the processing needs and the fact that most large Automotive companies have realized that eventually they need to have the capability to analyze the data on the cloud because if you look at Tesla one
Of their main competitive advantages is is that they get information from all the batteries from every user back every second and that kind of information gives you so so much insight into how you can design better batteries and you can’t just do that right like purely on
The edge that I think there’s common understanding that you need to have a proper cloudbased solution to analyze your battery data and so you might as well start that with your battery R&D test data in particular because it’s pretty large right we we’re talking about huge amounts of battery test data
That come off these huge facilities now and so yes it’s all cloudbased so what does the customer have to do how do you integrate into the process of your customer so let’s see if you wanted to basically just run a pilot then it’s on an Amazon web service Cloud into which
You whether you use an API or anything else just push your test data in production it depends a little bit on what kind of test stand equipment you use for example we really like working with national instrument who have a Enterprise architecture called system link which basically is something that
Has all of the test systems locked in and pushes all of the data essentially into some sort of Cloud solution so there’s a various like there’s a variety of test providers keyside National instrument who run these massive battery test centers and their job pretty much is and that’s why we like working with
Them is to get the data into some sort of cloud or SQL database and then we pick up the data from the web server and then help users figure out right like what the optimal test plan is and you mentioned data Gathering is not the problem in this approach right um no in
Batteries as you know I’ve worked in a lot of different engineering areas for the last five years and normally if you ask me what is the biggest problem I would immediately sh data gathering in batteries like we all Gathering battery data every day our laptops our phones
It’s quite cheap and easy in some sense to gather battery data because there sensors on them and you just need to go Curr in count out temperature controlled conditions are harder so it’s very easy to get a lot of battery data it’s much harder to then figure out hm what do I
Actually do with it so what are your next steps so you’re now in this battery Market maybe we have call in two years again are you still in the battery Market or do you see other markets so for me the vision behind monolith hasn’t changed at all for me batteries are just
An amazing example of where the whole story comes together really really nicely and so whenever I talk to any potential client or I will draw up a story from you design your product you simulate it you test it you manufacture it and you put it into market and all
Along this journey right there’s the same steps that Engineers take right like they have to figure out a test plan they have to come up with a model they have to figure out the root cause of specific issues that they’ve had with their product they need to validate that
Their data is true in these specific use cases they are completely identical whether I’m building a rocket an aircraft a shampoo bottle or a battery and I think for me this has been one of the biggest learnings at monolith of the last years is that there’s R&D of
Products has the same problem statements and pain points and that you can build machine learning solutions for each of those pain points and they really scale across the different areas and industries and so for me right now battery is just the one where there’s the most interest in the market and
Where we as a company are really motivated to just help it because of the climate change implications and the environmental impact and the fact that we need more ethical batteries like we’re very motivated to take batteries as a specific area and just push it forward as fast as we can but what I’m
Trying to say is we want to in the next let’s say year or two want to build out the perfect solution for batteries where a battery R&D team can from design to in the field can learn from all of the battery data sets along the way and
Solve all of the problems that can be solved with machine learning right that’s our Focus and then the same use cases we’ve already applied them to fuel cells hydrogen fuel cells we have a couple of hydrogen fuel cell examples so I don’t know what the next exciting
Thing will be I think batteries will be hot for a while but the same use cases and I’ve seen this again and again over the last five years they really apply across industry verticals and across different products and you find them again and again and the vision behind
Monolith is and remains building a platform that allows research and development engine Engineers to get access to these machine learning toolboxes to go and validate their data figure out the root cause optimize their test plan predict the result of a complex system and an unknown Vision come up with a model these different
Things they are pretty Universal Richard thank you very much it was a pleasure all the best and take care thank you Robert Oh