When AI meets physics, product development gets faster

Shorter lead times, rising material prices and uncertainty in global supply chains. Add new sustainability regulations and growing demands for production methods that use less material and energy. For manufacturing companies, adapting existing processes or introducing new ones can be both costly and time-consuming.

The PINNForm project – Physics-informed and AI-driven forming for resilient and sustainable manufacturing – addresses the need to develop and validate new manufacturing methods with fewer resources. Together with STILFOLD and AirForestry, RISE is exploring how AI and machine learning can be used to optimise the design and forming of metal components.

Unlike conventional data-driven AI approaches, physics-informed AI incorporates physical laws into the models, reducing their reliance on large amounts of training data.

“Ever since intelligent digital systems first emerged, there has been an ambition to use them to explore what is physically possible in the real world. Previously, this was an extremely complex and resource-intensive process. Today, however, physics-informed AI can suggest ways to optimise a metal forming process, for example by reducing material waste or improving energy efficiency,” explains Ramin Moshfegh, Senior Researcher and Business Developer at RISE.

Physics-informed AI accelerates the design process

For STILFOLD, which has developed a method for folding thin metal sheets into strong, lightweight structures with minimal material waste, the project offers an opportunity to develop design and optimisation tools that could shorten the time from initial concept to a working production solution.

STILFOLD's origami-inspired technology aims to reduce weight, material consumption, the need for specialised tooling and carbon emissions. By integrating the laws of physics into the design process, the need for large datasets and extensive testing can also be reduced. The AI model already incorporates knowledge of how metal bends, responds to loads and deforms. It can also predict the final shape of a component after the material springs back following forming.

“The aim is to simplify the digital design process so that we can create digital representations of curved, folded and complex structures more quickly. Machine learning allows us to generate more design alternatives, compare them and identify solutions we might not otherwise have considered. We can experiment with these digital representations to find the optimal design in terms of load-bearing capacity and weight,” explains Jonas Nyvang, CEO and co-founder of STILFOLD.

The project's second industry partner, AirForestry, has developed drone-based forestry technology that enables forest thinning using drones equipped with harvesting tools suspended from long cables. Within PINNForm, the focus is on the company's saw blade. This component is being designed using STILFOLD's origami-inspired technology within the physics-informed AI model.

“How much force does the blade need to withstand, and how much strength must it provide? How can we achieve this through geometry, manufacturing methods and material selection? So far, we have found that geometry – how the metal is shaped and folded – has an extremely important influence on the outcome. Much of our work focuses on enabling machine learning to account for these geometric effects, before adding the effects of material choices, manufacturing methods and other parameters,” says Jonas Nyvang.

For AirForestry, the project means that its existing product can be further developed in a digital environment before changes need to be tested physically. This makes it easier to compare different approaches and rule out unsuitable solutions before moving to physical testing.

Potential for greater resilience and flexibility

This type of development requires expertise across several specialised fields, from AI and advanced mathematics to materials science, manufacturing technology and robotics, as well as access to substantial computing power. Few smaller manufacturing companies have all these capabilities in-house.

In the project, RISE brings together the necessary resources and combines them with the industry partners' knowledge of their products and processes. This allows the companies to explore and validate new solutions without having to build up all the expertise and infrastructure themselves. RISE also provides access to MIMER, a national AI infrastructure located in Linköping.

“When we introduce more design parameters and advanced computational mechanics, the simulations quickly become extremely demanding and require enormous computing power. Being able to provide access to and use the national infrastructure through MIMER is essential for training these models efficiently. It is a tremendous asset for start-ups and small and medium-sized enterprises,” says Ramin Moshfegh.

So what is the potential of physics-informed AI models such as those being developed in PINNForm? According to Ramin Moshfegh, the technology could make Swedish manufacturing companies more resilient and flexible. Companies that need to develop new products faster or adapt their production processes – for example, to use more circular materials or recycled steel – could use the technology to test more ideas digitally before making costly investments. This would reduce both the risks involved and the time needed to move from concept to finished product.

“The idea is to make highly advanced technology accessible to companies with limited resources. This means smaller companies would not need to make enormous investments in specialised tooling. Our vision is for the software we are developing to take a solution directly from an office computer to the factory floor. Ultimately, it is about giving Swedish companies better opportunities to compete with companies in other European countries, as well as the US and China,” says Ramin Moshfegh.