Revolutionizing industries with AI-powered digital twins
IBM Research’s foundation models are powering simulated versions of complex systems, which promise to accelerate technological progress.
Digital twins have long been considered game-changers across industries. Whether in automotive and aerospace, or manufacturing, energy grids, and supply chain logistics, these virtual models of real-world systems have promised to transform operations through predictive modeling and real-time analytics. Until now, however, the technology has fallen short of its promise, due to limitations in modelling complex physical systems with sufficiently high accuracy.
IBM Research is working to change that. By applying foundation model techniques originally developed for language processing, our researchers created a framework for building AI-powered digital twins that are fully data-driven, self-improving, and capable of accurately predicting complex system behaviors. This advancement has broad implications, with applications spanning multiple industries that rely on high-fidelity simulations to optimize performance, improve safety, and reduce costs.
One of the most immediate and impactful applications of this approach is in electric vehicle (EV) battery development. EVs’ average range roughly tripled between 2010 and 2021, and their longevity is now comparable to that of internal combustion engine vehicles. But range anxiety and uncertainty around battery life remain major barriers to their wider adoption. Car buyers still worry about running out of power without access to a charger, and in the longer term, they’re concerned about battery packs degrading or failing prematurely. Despite recent improvements, it can take years to develop and test new batteries that last longer and enable cars to drive farther. Further progress is coming slowly for US and European automakers, especially compared to the relatively speedy development timelines from Chinese EV manufacturers.
Sphere Energy, based in Augsburg, Germany, is collaborating with IBM Research to specialize and validate IBM’s foundation model-based digital twins to accelerate EV battery development and testing. By utilizing AI-powered simulations, manufacturers can replace years of physical testing with highly accurate virtual models that predict battery performance and degradation under real-world conditions.
At the heart of the digital twin innovation are AI model architectures and techniques popularly used for language foundation models – the ones seen in today’s large language models. But whereas LLMs are trained on text datasets and deployed on text-based tasks, these models are designed to handle the complex chemistry and operations of battery components.
In this way, IBM Research and Sphere are collaborating to train foundation model-based digital twins of EV batteries, which battery manufacturers can put through their paces to gather accurate predictions of performance, safety, and thermal behaviors. By predicting hundreds of cycles based on as few as 50 initial cycles, foundation model-based digital twins significantly increase testing throughput compared to the traditional approach, where EV manufacturers would need to physically develop new battery technologies. This advancement is estimated to save millions of dollars and years of road testing, according to Sphere.
“Battery engineering is based on data, not language, so we’re using foundation models to simulate the next best data point instead of the next best word,” said Sphere co-founder Lukas Lutz. And the test results speak for themselves: Sphere can predict within a 1% margin of error how a battery will behave and age when performance starts to drop in a nonlinear way. “This was simply not possible before,” said Lutz. “Current simulations assume a linear decay and therefore are not able to predict the important nonlinear regime.”
With foundation model-based digital twins, Sphere intends to disrupt the lengthy battery development cycle. When an original battery equipment manufacturer wants to make new battery cell technology to sell to automakers, they must validate it for several years in a car. This industry-standard lifecycle testing isn’t a regulatory requirement, but it’s essential for seeing how a battery cell ages in hot or cold climates, under gentle and aggressive driving, with fast or slow charging — real world conditions that affect how quickly a battery ages.
Rather than road testing for years, battery makers can deploy the foundation model-based digital twin that simulates a huge range of driving conditions and models how they’ll affect batteries.
At Sphere’s battery testing facility in Augsburg, more than 1,500 cells of various sizes, formats, and chemistries run 24/7 to provide the data used for AI model pre-training. To date, the model was trained with data from more than 4,000 tests. This data is fed into the IBM foundation model architecture developed by the IBM Research team to learn the state of a specific device and generate predictions for various descriptors. In the case of batteries, these descriptors include voltage, current, and capacity. Specifically, the digital twin model relies on an encoder-decoder architecture based on vanilla transformer blocks that are extended to integrate numerical representations. When it was pre-trained on Sphere’s broad datasets — which also include metadata about the cell chemistry — the model learned the transitions between different states of a system.
The foundation model’s outputs can be generalized across a variety of use cases with minimal fine-tuning, which is a big deal in the battery industry because it allows manufacturers to use the same model for predicting a battery’s behaviors across various vehicles, requiring only limited testing on a few edge cases to validate the simulations.
“Foundation model-based digital twins construct a comprehensive digital representation of a system,” said Teodoro Laino, distinguished research scientist at IBM Research. “This enables the extraction of multiple interrelated measurements, providing deeper insights and a more holistic understanding of performance.”
“Sphere’s decision to work with IBM was driven by the uniqueness of our foundation model approach for digital twins, which is a distinctive capability developed at IBM,” Laino added.
But batteries are just the beginning. The transition from traditional simulations to real-time, predictive modeling signals a broader shift in how industries can use foundation models to drive efficiencies and accelerate innovation. Whether in battery development, manufacturing, energy grid management, or predictive maintenance for industrial systems, data-driven digital twins can reshape how businesses approach complex system challenges.
Foundation models have the ability to capture intricate relationships, and as such, foundation model-based digital twins are well suited to predicting battery aging. They can take into account the battery’s internal state, environmental conditions, and application scenarios — leading to more precise forecasts and performance optimization under a wide range of conditions. This flexibility allows for highly accurate predictions of aging and performance with minimal additional data input for fine-tuning on client-specific data, resulting in an estimated 50% reduction in development time and cost for new batteries.
Digital twins do more than speed up the testing process, said Lutz. They also make it possible to perform steps in parallel that usually need to happen in sequence. For example, a battery manufacturer can share digital twin modeling data with other business units earlier on in a project, enabling them to incorporate new cell technology faster. “When a cell is in the validation phase for years, it’s hard to quickly incorporate a new anode or cathode technology — which is happening every 6 months by the way,” Lutz said.
So even though new battery breakthroughs have been happening quickly in the field, cell testing takes so long that new innovations simply can’t be incorporated because it doesn’t fit into the testing cycle. “We want to change that,” said Lutz.
Improving these parallel processes could start to give Western EV manufacturers a competitive edge in the EV market. Chinese manufacturers have pushed down vehicle development times by making major investments in battery technology and fostering cooperation across the auto manufacturing industry.
The new trained models developed in this collaboration are already in use with battery manufacturers, car manufacturers, and other related players such as power toolers in Europe, demonstrating how AI-driven simulations can transform the industry.
As this technology advances, IBM envisions AI-powered digital twins playing a key role not only in optimizing performance, safety, and sustainability of batteries, but also in improving broader industrial applications. AI-powered digital twins have the potential to drive smarter, safer, and more efficient industrial processes across a wide range of sectors, including predictive maintenance in manufacturing, optimizing energy grids, improving supply chain efficiency, and enhancing aerospace system reliability.