Project Description

Key information
Project in the Spotlight: T18019
Market: Advanced Metals

Written M2i Program Manager: Viktoria Savran

At first glance, it might seem a simple matter: fill a blast furnace with iron-bearing materials and let the chemistry take over. But in reality, the way you load the furnace can make or break the efficiency of iron production. That’s because the materials used—pellets and sinter, of different shapes and sizes—don’t just sit still. They segregate. They shift. And they build up in unpredictable ways, altering the internal structure of the burden and ultimately impacting permeability, gas flow, efficiency and productivity.

For years, this complexity posed a challenge to engineers and operators at steel producers, like Tata Steel. And while physical experiments offer some insight, they’re expensive, slow, and can’t track the particle-scale dynamics that drive these effects. That’s where discrete element modelling (DEM) comes in.

And more recently, that’s where PhD researcher Ahmed Hadi, working at TU Delft under the guidance of Prof. Dingena Schott and with close collaboration with Allert Adema from Tata Steel, stepped in to take things to the next level.

The DEM-OC project focused on one of the steel industry’s biggest headaches: how to model, predict, and ultimately control segregation of iron ore pellets and sinter during blast furnace charging. The problem is that when these components flow through chutes and hoppers, they don’t mix evenly

Instead, the final heap in the receiving bin often shows significant radial segregation, with pellets or sinter clustering more in certain areas. This can cause uneven melting and gas flow in the furnace. While DEM provides a virtual laboratory for exploring these dynamics, it comes at a steep computational cost.

To calibrate a DEM model accurately, hundreds of simulations might be needed, each representing slight variations in friction coefficients, restitution parameters, or other contact properties between particles and walls. With multi-component mixtures, this process gets exponentially more complex. Even defining the correct initial configurations (ICs) of material in the hopper—for instance, layered, mixed, or partially segregated—introduces new challenges.

This is where the DEM-OC project introduced a game-changing innovation: the development of machine-learning-based surrogate models. Instead of running thousands of DEM simulations, Ahmed Hadi trained surrogate models to approximate the relationship between DEM input parameters and the resulting segregation patterns. And to do this efficiently, the team introduced transfer learning.

“We wanted to build a model that could not only perform well on known scenarios but adapt quickly to new ones with minimal extra data,” explains Hadi. The idea was to train a surrogate model on several known initial configurations, then transfer this knowledge to unseen ones. With just a handful of new simulations—sometimes as few as five—the model could adapt to predict segregation behaviour with high accuracy.

To build the training datasets, the team used a cost-effective design of experiments strategy known as Definitive Screening Design (DSD), which allowed them to efficiently vary 15 DEM interaction parameters across five distinct ICs. These included friction coefficients, restitution factors, and other variables critical to particle-particle and particle-wall interactions.

Once the data was collected, a wide range of machine learning techniques were tested: from support vector machines to neural networks. But the standout performer was Gaussian Process Regression (GPR), which provided both accuracy and speed, especially when paired with Bayesian optimisation for tuning model hyperparameters.

For Tata Steel, the implications are significant. Being able to reliably predict segregation outcomes from charging strategies means better control over burden distribution, improved gas flow, and more efficient steelmaking. As Allert Adema of Tata Steel puts it: “This work gives us a powerful new lens into what happens inside the blast furnace before the first spark flies. We can now foresee and control the internal structure of the burden in ways that were simply not possible before.”

For academia, the DEM-OC project showcases the potential of combining traditional physical modelling with cutting-edge AI tools. “We’ve always known DEM could model these systems in detail,” says Prof. Dingena Schott. “But now, we can bridge the gap between theory and industrial reality. With machine learning, we’re cutting down the time and computation needed to make DEM practical for large-scale, real-time applications.”

And for the researcher himself, the journey has been transformative. “This project taught me how to connect physics, data, and industrial need in a single framework,” reflects Hadi. “Working so closely with Tata Steel gave me a clear sense of how our research can drive operational impact. That’s the most satisfying part.”

The DEM-OC project (T18019) received funding through M2i’s PPS instrument and was carried out in collaboration with TU Delft and Tata Steel. It marks a strong step forward in bringing AI-enhanced modelling into the heart of traditional industries like steelmaking.

As M2i continues to catalyse collaboration between academia and industry, projects like this one highlight a powerful message: even the most established, high-temperature processes can benefit from smart, cutting-edge algorithms. For those interested in exploring similar innovation pathways, M2i remains a ready partner in enabling smarter, faster, and greener materials solutions.