University of Birmingham, EvoPhase and Simcenter M-STAR recognised for AI-driven optimisation of sustainable biopolymer manufacturing
The University of Birmingham, in collaboration with Simcenter M-STAR and EvoPhase, was granted access to 3500 GPU nodes on the Leonardo to develop a digital platform for the optimised production of sustainable biopolymer products using AI-driven design exploration. At the SC 26 conference held in Chicago, they were awarded the HPC Wire Reader's Choice Award for Best use of HPC/AI in Industry (Automotive, Aerospace, Manufacturing, Chemical) for their work on the multidimensional optimisation of twin screw extruder systems.
Support from ENCCS and Sweden AI Factory
The project was supported by ENCCS and the Sweden AI Factory, which provided guidance on access to and effective use of large-scale GPU computing resources. Through this support, the team was able to fully leverage the EuroHPC Leonardo system, helping accelerate the development and execution of the large-scale optimisation workflow behind the award-winning results.
AI-Driven Optimisation of Twin-Screw Extruders for Sustainable Biopolymer Manufacturing
Organisations involved
The University of Birmingham is a leading research-intensive institution in the UK, with strong expertise in chemical engineering and formulation engineering, heavily applied to industrial systems. EvoPhase is a UK company specialising in the development and application of AI-driven optimisation methods; their HARPPP workflow facilitates the physics-informed autonomous redesign and performance improvement of a diverse range of industrial systems. Simcenter M-STAR is a simulation software company in the USA which offers a GPU-native CFD solver built on the Lattice Boltzmann Method, purpose-built for simulating complex transient flows involving liquids, gases, particles, heat transfer and multiphase phenomena. Its GPU-accelerated architecture delivers simulation speeds orders of magnitude faster than traditional finite-volume CFD, enabling high-fidelity results in hours rather than days.
Technical challenge
Only 9% of all plastics ever produced have been actively recycled, with the remaining 91% destined for landfill or incineration. While biopolymers offer a promising solution, their industrial-scale production via twin-screw extrusion remains extremely challenging. Twin-screw extruders are "black box" systems.
Their internal process dynamics are complex, and traditional experimental and CAD-based design approaches are too slow and resource-intensive to explore the full design space. This bottleneck severely limits the discovery of high-performing, non-intuitive extruder geometries
Proposed solution
By leveraging Leonardo's GPU cluster alongside the GPU-accelerated Lattice Boltzmann solver Simcenter M-STAR, a novel, fully parametric geometry framework was coupled with HARPPP. This integrated digital platform autonomously generates, simulates, and evaluates twin-screw extruder designs at unprecedented scale, achieving physics-informed exploration of over 1,000 novel non-intuitive candidate designs in under five days.
This allowed the discovery of an innovative design with a 3-fold improvement in mixing performance and over 57% reduction in power draw, relative to the industrial standard, hence a step change in extrusion process design beyond the conventional.
Business impact
Access to Leonardo's approximately 3,500 GPUs enabled the full optimisation workflow to run at maximum capacity, something impossible on conventional computing infrastructure. For each design, it took 3 hours in total from the mesh generation stage to simulation and performance ranking, allowing the testing of 1600 novel non-intuitive designs in just over 7 days.
Given that the optimisation is physics-informed, it learns the individual and combinatorial impact of all geometric properties on the performance metrics of interest, i.e. mixing and shear, at an accelerated rate, leading to the discovery of unconventional designs that outperform the industrial standard.
The cost of this optimisation is electricity, in comparison to traditional and experimental approaches that are time-consuming, waste valuable raw materials, require long design manufacturing times, are energy intensive and are limited in design exploration rate. Therefore, the digital platform stands to transform how manufacturers develop and scale extrusion processes for biopolymer products, reducing development time and cost while enabling discovery of designs that significantly outperform industry standards.
Key results
- Achieved a design exploration rate of over 1,000 novel extruder geometries in under 5 days, far beyond the reach of traditional methods.
- Discovered non-intuitive designs delivering a 3-fold improvement in mixing performance and over 57% reduction in power draw.
- Physics-informed designs are ready to integrate directly onto existing physical twin-screw systems, ensuring immediate industrial applicability.