Polymer engineer with a vision for digitalization

I started my journey with plastics in 2016, and worked my way up from managing extrusion lines, to compound formulations and scale-ups, to new material R&D. It did not take long for me to realize the inefficiencies in the day-to-day development of plastics and visualize the opportunity for utilizing machine learning in this specific industry.

Data-driven methods and artificial intelligence can transform critical workflows, in an industry where eliminating waste and downtimes can mean keep the lights on. I first adopted machine learning in formulation optimization and constructing structure-property correlations, reducing development times and reducing the number of extruder trials by 60%.

The real cost saving comes from ensuring proper utilization of machinery and eliminating downtime from predictable failures. Utilizing computer vision, I translated routine machine readings to data that can be analyzed by AI which can predict material failures well before any machine trial is taken.

I spent considerable time with medical device product development with Becton Dickinson and Company (BD), helping project leaders identify the right materials, learn from product failures and pinpoint the root cause. I used large language models to accurately translate product requirements to material property requirements, reducing product development timelines by >33% and simultaneously evaluate multiple materials, without conducting extensive experiments.

Where AI can fail, and where I have failed in the past?

Using AI with polymers is not as straight forward as in other industries. For the past 5 years, I have faced more failure and scepticism than I can count, and that's exactly why I want to help other to avoid the same pitfalls. Building the model is relatively the easy part, the challenge lies in knowing which question to point it at, and in being willing to say when the data cannot answer it. A model that predicts tensile strength to within 3 MPa is useless if tensile strength was never the property that decided whether the part passed.

I derive my knowledge from both real-world industrial problem, and published literature which often contain synthetic data. The key components for a successful outcome are well structured data and well understood parameters, both of which can be hard to find in the real world setting. Papers report R² on held-out sets drawn from the same narrow chemical space as the training data, and quietly omit that the model collapses the moment it meets a filler it has not seen. I try to balance between published knowledge and real world challenges to solve problems, working closely with clients and delivering realistic results.

Working together

I work with a small number of clients at a time, usually compounders, moulders and processors who have a specific expensive problem rather than a general interest in AI. If that is you, the fastest way to find out whether I can help is to tell me what you have already tried.