When we use AI to model and predict materials innovations, we run into a problem – it is hard to realistically model complex sequences of events. It turns out that this problem is similar to one found in other fields where we try to replicate the real world, such as video generation or autonomous driving. An AI technology that is becoming a lynchpin in these arenas could have a big impact in materials research. Intellegens is leading a Frontier AI project to investigate how world models could drive optimized materials and processes. In this month’s blog, we dip into world models and learn how they could help materials science.
Early attempts at AI video generation worked by analyzing and learning from the frames of source videos. The method would understand these frames simply as a set of pixels and try to identify and anticipate visual transitions to create new frames. But such videos struggled to look like realistic renderings of the world. They didn’t incorporate a true understanding of the full history of the video sequence, or why the objects in the video behaved the way they did. For example, they often failed to make sensible decisions about what should happen to elements that passed temporarily out of view.
World models were a solution to this problem. A world model learns from all of the available data to construct a representation of system dynamics, creating a set of rules to mimic physical reality, rather than relying on surface-level correlations between successive frames. AI video systems achieved a leap in realism, stability, and continuity.
Advanced manufacturing processes are also fundamentally defined by sequence, duration, and history. The properties of an engineered alloy or a high-performance cutting tool are shaped not merely by the elements from which they are made, but by a precise chronological journey: a delicate balance of heat treatments, mechanics of mixing, and thermodynamic transitions across time. Yet, conventional machine learning predictions are largely blind to this depth. They reduce rich, path-dependent manufacturing histories to static, two-dimensional snapshots of input parameters and final outcomes.
As in the world of video generation, world models could change the game. In our Frontier AI project, we’re constructing generative world models that have been trained using a combination of empirical data and data from physical simulations, providing the models with a grounding in physical laws. This enables the algorithm to understand the intrinsic mechanics of processing. Rather than navigating experimental landscapes through blind trial and error, these models can then simulate entire synthesis routes, evaluating multi-step pathways before a single furnace is heated or a single batch is mixed.
Our initial investigations focus on a classic challenge in materials engineering: understanding how the operational performance of cemented carbide drill bits emerges from complex microstructural evolution during sintering and brazing. Our framework uses high-fidelity computational simulations based on open-source tools such as CALPHAD and Kawin to model out-of-equilibrium thermodynamics and precipitation kinetics.
We’re coupling this with some smart state-of-the-art AI. Recurrent state-space architectures provide a model type designed for sequential data, while adaptive training targeting multiple objectives simultaneously means that the AI learns flexibly and constructs a model that can usefully guide decision-making while inherently quantifying its uncertainty.
The models are being developed and validated in collaboration with a world-leading manufacturing organization, with support from Innovate UK.
By bridging the gap between physics-based simulation and data-driven generative models, this framework offers a glimpse into a future where physical experimentation is guided by immediate predictive feedback. As these latent spaces mature, world models promise to fundamentally reshape materials manufacturing, transforming multi-step development from an exercise in empirical trial and error into a deliberate, accelerated journey of physical discovery.
Technology that is today enabling video games to emulate the real world could turn full-circle and enable real-world research that delivers the next blockbuster material.
