polySCOUT

A new polymer typically takes twenty to thirty years to move from concept to market, and predicting how a material will behave remains difficult. This practice designs directly from the specification of the intended product. Machine-learning models predict polymer behaviour across technical performance, while chemistry-based approaches consider safety, sustainability and cost effectiveness at the same time, so that material choices can be made early in development. The main outcome of this workflow is a molecular design fitting with a set of desired functional properties. The models draw on a curated data foundation that combines the organisation’s own research results, technical data sheets for commercially available materials, and data from the scientific literature selected by polymer specialists. The foundation grows and the models continue to learn from it. After expert review, designed materials are synthesised, characterised and validated in the organisation’s laboratories, and where required scaled into plastics through an industrial partner network.

The programme was established in 2022 within a national applied research organisation, bringing together its research groups in data science and materials science from different parts of the organisation. The programme responds to three obstacles in the field. Data on polymers and their properties is scarce and fragmented. Existing machine-learning representations of polymers, known as fingerprints, are oversimplified: they omit essential polymer characteristics, especially for copolymers (polymers consisting of multiple “types” of polymers). Additionally, existing models design for functionality alone, while new materials are required to be safe and sustainable by design and to have appropriate end-of-life options such as recyclability or biodegradation.

Modelling and laboratory work sit in the same organisation. Industry engagement runs through joint development agreements and contract research, with electronics, automotive and textiles identified as target sectors. With the spin-off into a start-up company further commercialization of the technology will be enabled while maintaining close ties to the resources available at the research organization.

Compute resources were managed through internal infrastructure and Azure services. As data quantities in the polymer space are limited, compute demand was rarely larger than a single H100. Developed models are a combination of pre-trained ModernBert models finetuned on internal data, and proprietary predictive neural networks that take the Transformer embeddings and predict properties.

Within the initiative, a tool (interface) was developed that integrates the various models and workflows to allow users to run through the typical AI-accelerated design workflow by themselves. However, outputs of this workflow should ideally be reviewed by a polymer scientist before the next step of synthesizing the suggested polymer is taken.

The developed methods provide suggested molecular designs of (co)polymers fitting with a set of desired physical properties. This allows polymer researchers to more efficiently explore the design space of novel (bio)polymers to help replace existing fossil-based plastics. Where previously a scientist could evaluate a dozen options in hours, the developed models can evaluate thousands within seconds.

As the methods are being spun out of the research organization into a start-up company, all data, code, and models are proprietary. Internally, these are documented and managed through git repositories. The tool is openly advertised as AI-based and it is clearly communicated how results are generated.

AI was needed for this tool as conventional methods for property prediction are in many cases slow or unsuitable for polymer materials. AI was seen as a novel method to fit models directly to data and speed up predictions to allow for the desired rapid exploration of the molecular design space. Resource consumption was fairly limited and fit to the scale of the available data.

Since the developed methods make are used to make predictions outside their training domain, trusting outcomes is a challenge. For this reasons, methods for quantifying uncertainty in the model predictions and comparisons between new predictions and known training samples were introduced. In addition, polymer scientists should always be involved when defining inputs and evaluating outcomes of the tool.

https://www.tno.nl/en/sustainable/industry/sustainable-materials/circular-plastics/biopolymer-design/https://ventures.tno.nl/portfolio/polyscout/