跳到正文
Unite.AI·· 3 小時前AI 評分34

Holyvolt 創辦人兼 CEO Mathias Ingvarsson 訪談

Mathias Ingvarsson, Founder and CEO of Holyvolt – Interview Series

AI 導讀

Holyvolt 創辦人兼 CEO Mathias Ingvarsson 討論公司如何結合機器學習與實驗,推進電池材料開發及由預測走向驗證和規模化生產。Holyvolt 收購 Wildcat Discovery Technologies,並利用其涵蓋超過 500,000 個電池、歷時 18 年及 165 million 次循環的實驗數據建立 AI 電池研發平台。

正文

Mathias Ingvarsson is Founder and CEO of Holyvolt, which he established in 2022 to pursue new approaches to battery technology. His career spans Tempur-Pedic’s international expansion and entrepreneurial ventures across several industries. At Holyvolt, he focuses on connecting scientific expertise, experimental data, and manufacturing capabilities to accelerate the development of battery materials, including through the acquisition of Wildcat Discovery Technologies.

Holyvolt develops energy production and storage technologies, combining high-throughput materials discovery with precision screen-printing approaches to manufacturing. Through Wildcat Discovery Technologies, its battery research platform connects laboratory experimentation, historical testing data, and AI-enabled materials development. In this interview, Ingvarsson discusses how scientists use machine learning alongside physical experiments, why failed formulations can be valuable data, and what it takes to move from faster predictions to validated materials and scalable production.

Your career spans Tempur-Pedic’s international expansion and several entrepreneurial ventures. What led you to found Holyvolt in 2022, and what opportunity in battery technology convinced you to build a company in such a technically demanding industry?

I’ve built businesses in everything from mattresses to underwear, so batteries might seem like an unusual next step. But I’ve always been drawn to industries where I see an opportunity others are missing.

I knew very little about battery chemistry when I started Holyvolt, and I actually think that helped. Sometimes not knowing all the reasons something can’t be done is a superpower. Jensen Huang once said that if he’d known how hard building Nvidia would be, he probably wouldn’t have started. I can relate to that.

What I saw was a Western battery industry trying to beat China at a game China had already won. The industry was investing billions trying to replicate technology China had spent decades perfecting. I couldn’t understand why we weren’t putting that same energy into developing something better. Batteries are going to power almost everything around us, and I saw an enormous opportunity to rethink how we develop them. That’s what led me to start Holyvolt.

When you evaluated acquiring Wildcat, what convinced you that its experimental history could become the foundation for an AI discovery platform? What did you need to establish about the data beyond its sheer volume?

The sheer amount of data immediately caught my attention. More than 500,000 battery cells tested over 18 years and 165 million cycles. And the more we looked, the more impressed we became with the rigor of the experiments and the breakthroughs Wildcat had already achieved. That kind of data, collected consistently over decades, is incredibly difficult to replicate.

I’m not a battery scientist, so I brought in people who knew far more than I did. Experts in batteries, AI, and materials science came to see the lab, and frankly, they were blown away.

Then you look at the companies coming to Wildcat, including BMW and other world-renowned companies across the entire value chain. These are companies with enormous R&D budgets and some of the best scientists in the world, and they were turning to Wildcat for help. That was the ultimate validation for me. I became convinced that Wildcat had built the most sought-after battery discovery platform in the world. And with AI, I believe we’re only scratching the surface of what it can do.

Failed experiments are a central part of your data advantage. Can you share an example of what a model can learn from an unsuccessful battery formulation, and how your team distinguishes a meaningful chemical failure from a problem with how an experiment was conducted?

Failure is part of chemistry. That’s how we’ve made scientific progress for hundreds of years. You try something, it doesn’t work, you learn from it, and you try again. AI needs to learn the same way.

We’ve spent 18 years running high-throughput battery experiments, so our data includes everything from failed formulations to materials we’ve successfully optimized. That’s incredibly valuable when training a model.

Take lithium-metal batteries. Unlike traditional graphite batteries, they can experience a sudden drop in capacity that’s very difficult to predict. We’ve developed machine learning models using real battery cycling data across different electrolyte formulations, electrode designs, and testing protocols to predict when that drop might happen.

Our initial models can predict these drops with reasonable confidence, and we’re designing more targeted experiments and testing batteries under different conditions to refine those predictions.

Of course, that data is only valuable if we can trust the results. We’ve spent years refining our laboratory processes to make experiments highly reproducible and minimize errors. We also run multiple instances of the same experiment, so we can tell whether a poor result reflects a genuine problem with the material or an anomaly in the testing process. That way, we’re not teaching our models the wrong lessons from a single unrepresentative result.

Could you walk us through a discovery cycle, from a model proposing a material to the laboratory testing it and incorporating the results? Which decisions does the AI make, and where does a scientist’s judgment remain essential?

It starts with our scientists. They design an experiment based on what they’re trying to learn, and our high-throughput platform runs the tests. Our machine learning engineers then use that initial data to build a model and see whether it can predict the results of future experiments.

Then we test those predictions in the lab. If the results match, we have more confidence in the model and can use it to evaluate new formulations. If they don’t, our scientists investigate why. Sometimes the model needs more data, which means designing entirely new experiments or testing under different conditions. We feed those results back into the model, refine it, and repeat the process.

There’s a lot of excitement about AI independently discovering new materials, but we’re careful about what we claim. At this stage, our scientists are still designing the experiments and making the key decisions. AI helps us make predictions, while our laboratory tells us whether those predictions actually hold up. By combining our scientists’ expertise with AI, we’ve already been able to dramatically reduce the time needed for certain experimental design tasks, from weeks to just minutes.

A historical dataset reflects the materials researchers previously chose to investigate. How do you prevent your models from staying within those familiar areas, and determine when an unfamiliar chemistry is promising enough to justify physical testing?

That’s exactly the right question, and it’s why I believe our laboratory is even more valuable than our dataset.

A model learns from the data you give it, so naturally it’s going to be more reliable when predicting things it’s already seen. The real opportunity is using AI to explore entirely new areas of chemistry, where that historical data may not exist.

That’s where our high-throughput platform gives us an enormous advantage. We can rapidly generate new, reliable experimental data across hundreds of formulations, giving our models a foundation to learn from even in unfamiliar areas of chemistry.

And because we can test predictions in the lab, feed the results back into our models, and repeat the process, we can continuously improve their accuracy. Every experiment makes the next prediction more informed.

I believe that’s our biggest competitive advantage. We have 18 years of historical data, but we’re not limited by it. We have the ability to generate the new data we need to explore chemistries that haven’t been thoroughly investigated before.

How do you measure the improvement attributable specifically to AI, beyond what automated, high-throughput experimentation already delivers? What comparisons show that the models are finding better candidates or reducing the experiments needed to reach a useful result?

That’s actually a difficult question, because scientific discovery is a team effort.

Think about a group of scientists brainstorming a problem. One person might eventually come up with the solution, but that idea could have come from hours of discussion with everyone else in the room. Who gets the credit?

I think we should look at AI in a similar way. It’s another member of the research team. It can help scientists find relevant information faster, recognize patterns in experimental data, and understand a problem more deeply.

Our high-throughput platform already gives us a tremendous advantage in how quickly we can run experiments. We’re now working to understand how much further AI can take us. I don’t want to throw out a number we can’t properly support. What matters is whether we’re helping our scientists reach better answers with less time and fewer experiments.

Battery development involves competing objectives, including energy density, cycle life, safety, cost, and manufacturability. How do you translate those trade-offs into goals for the AI, and can changing the intended application lead it toward fundamentally different materials?

It really depends on what the battery is going to be used for. A battery for a drone has very different requirements from one going into an electric vehicle. You might prioritize energy density in one application and cycle life or cost in another.

We can define those targets quite clearly. Energy density, cycle life, safety, and cost can all be measured, so we can give the model specific performance goals based on the application.

The difficult part is working backward from those targets to a material that can actually deliver them.

Changing the application can absolutely change which materials we investigate. But our scientists still have to determine which chemistries are promising, design the experiments, and test whether those materials can meet the performance targets.

Holyvolt is advancing nickel- and cobalt-free disordered rocksalt cathode materials toward commercialization. What role has AI played in that development, and which remaining hurdles require manufacturing or engineering solutions beyond materials discovery?

DRX is a great example of why our approach matters. There are an enormous number of possible combinations of materials, and testing them one by one would take years.

At Holyvolt, our high-throughput platform allows us to rapidly test hundreds of different formulations and understand how changes in chemistry affect energy density, voltage, battery life, and performance. We’ve been able to identify promising materials much faster than traditional research methods would allow.

Now we’re combining that experimental data with machine learning to identify patterns and help our scientists decide which formulations to explore next. The more we test, the more we learn, and the better we can guide future experiments.

We’re already making progress toward commercial-scale production, and one of our biggest advantages is that we’re using well-established manufacturing processes. We don’t need to develop entirely new equipment or reinvent how these materials are produced. Our focus is on scaling production with our partners while maintaining consistent performance and competitive costs.

You have argued that proprietary scientific data can be a stronger competitive advantage than the models themselves. What makes that advantage compound over time, and how do you balance learning across customer projects with protecting each customer’s confidential research and intellectual property?

Some of the most powerful AI models available today are open source, and they’re only getting better. But you can’t simply download 18 years of battery research. That’s what makes our data so valuable.

At Holyvolt, we’ve tested more than 500,000 battery cells across more than 165 million cycles. We know which materials were used, how they were processed, how they performed, and where they failed. And because we’ve followed consistent testing methods over all these years, we can actually compare those results. That’s incredibly difficult for anyone else to replicate.

Every experiment teaches us something. That knowledge helps us improve our models, make better predictions, and decide what to test next. And the more experiments we run, the more valuable that foundation becomes.

We also work with some of the world’s leading companies, and they trust us with sensitive research. We take that responsibility seriously and have strong protections in place to safeguard confidential information and intellectual property.

As you expand the platform, what is the biggest constraint on faster discovery today: model capability, experimental capacity, the time required to validate battery life, or something else? What measurable milestone over the next two years would demonstrate that AI is changing the economics of battery development?

I think the biggest bottleneck today is the time scientists spend designing experiments and analyzing the results.

Think about the traditional process. A scientist asks a question, designs an experiment, runs the tests, analyzes the data, and then decides what to try next. We’ve already automated much of the physical testing through our high-throughput platform, but the thinking between those steps still takes a tremendous amount of time.

We recently applied a relatively simple AI model to part of the experimental design process. Something that would normally take six weeks took eight minutes. That’s a pretty extraordinary result, and it shows you what’s possible.

But speeding up experimental design is only the beginning. As AI helps us identify what to test faster, we also need to increase how quickly we can run those experiments. That’s why we’re planning to expand our laboratory capacity and invest in further automation.

Over the next two years, I want to see us dramatically accelerate the entire discovery process, from the first idea to a validated material. If we can combine faster decision-making with the ability to run more experiments, we can explore new chemistries at a pace that simply wasn’t possible before. That’s how AI can fundamentally change the economics of battery development.

Thank you for the great interview, readers who wish to learn more about battery materials, high-throughput experimentation, and AI-enabled energy technology development should visit Holyvolt.

來源:Unite.AI · unite.ai