AkzoNobel
Introduction
AkzoNobel, a global leader in paints and coatings, relies on Nuclear Magnetic Resonance (NMR) spectroscopy to analyse chemical compounds in its products. This technique is crucial for quality control, research, and innovation, but the manual interpretation of NMR spectra is a time-consuming process that demands significant expertise.
Recognising the need for automation, AkzoNobel partnered with the National Innovation Centre for Data (NICD) to explore machine learning-based clustering techniques that could improve efficiency and support knowledge transfer within the organisation.
The challenge
The process of analysing NMR data presents multiple challenges:
- 1. Time-Intensive Analysis: Experts manually interpret spectral data by recognising patterns based on experience. This approach, while effective, is highly dependent on a small number of specialists and is difficult to scale.
- 2. Knowledge Retention & Training: AkzoNobel has decades of archived NMR data, but there was no structured way to leverage it for training new scientists. As experienced chemists retire, the loss of institutional knowledge poses a risk to continuity.
- 3. Lack of Existing Software Solutions: Off-the-shelf software solutions for NMR clustering do not exist, particularly for the complex, multi-component mixtures involved in paint chemistry.
As Simon Welsh, NMR specialist at AkzoNobel, explained:
"With over 20 years of archived data, we knew there had to be a way to organise it better. But we weren’t sure if machine learning could actually handle the complexity of our spectra."
Jennifer Longyear, Research Biologist at AkzoNobel added:
"We realised that training new people in NMR analysis was incredibly slow because they lacked exposure to all the different spectra we had encountered over the years. We needed a way to help them recognise patterns faster."
The approach
NICD and AkzoNobel collaborated to build a data-driven solution for clustering NMR spectra. The project was designed as both a research initiative and a skill-transfer opportunity, ensuring that AkzoNobel’s team would be equipped to refine and extend the tool independently.
Data science methodology
To develop an effective clustering tool, the following technical approach was taken:
- 1. Data Processing & Feature Extraction:
- Each NMR spectrum consists of thousands of data points representing chemical shift and intensity values.
- To make clustering computationally feasible, the team used embedding techniques to reduce dimensionality while preserving meaningful patterns.
- 2. Deep Learning Models:
- Convolutional Neural Network (CNN) Autoencoder: Used to compress and reconstruct NMR spectra, capturing essential features in a compact representation.
- Vision Transformer (ViT) Autoencoder: Leveraged transformer-based architectures typically used in image analysis to process spectral data.
- Time-Series Feature Extractor: Captured statistical characteristics of NMR spectra, such as autocorrelation, variance, and peak prominence.
- 3. Clustering Techniques:
- UMAP (Uniform Manifold Approximation and Projection): Reduced high-dimensional data to a more manageable form while preserving structural relationships.
- HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise): Identified clusters based on density, effectively grouping similar spectra together.
By combining these methods, the team was able to create a proof-of-concept system capable of clustering NMR spectra in a way that aligned with expert human analysis.
The outcome
The collaboration led to several key achievements:
Validated Clustering Model:- The proof-of-concept successfully grouped spectra into meaningful clusters, helping chemists quickly identify known patterns and detect anomalies.
- Simon Welsh verified that the automated clustering aligned with his own expert assessments, confirming the system’s reliability.
- The AkzoNobel team gained hands-on experience with Python programming, machine learning fundamentals, version control, and experiment tracking.
- These skills enabled them to modify and extend the clustering tool beyond the initial project scope.
- The company has since begun developing a Shiny dashboard to provide an intuitive interface for searching and interpreting clustered NMR spectra.
- Future plans include refining cluster labels and integrating the tool into routine workflows.
“Working with Jennifer and Simon was a fantastic experience. They quickly picked up new skills and applied them creatively. The clustering approach we developed has real value for AkzoNobel’s analytical workflows.”
Dr Antonia Kontaratou, Data Scientist, National Innovation Centre for Data
Business impact and future prospects
Beyond the immediate technical benefits, the project has broader implications for AkzoNobel:
- Increased Efficiency: The tool reduces the manual workload on specialists, freeing them to focus on complex problem-solving.
- Faster Training for New Scientists: By providing structured, example-based learning, the tool accelerates the development of expertise in NMR interpretation.
- Data-Driven Decision-Making: The ability to mine historical data enables more informed research and development strategies.
“This project has fundamentally changed how we think about data. We’ve gone from manually recognising patterns to using AI to assist us, and that’s just the beginning”
Jennifer Longyear, Research Biologist, AkzoNobel
Conclusion
The collaboration between AkzoNobel and NICD demonstrates the power of machine learning in industrial chemistry. By transforming an expert-driven manual process into a scalable AI-assisted workflow, the project has set the stage for continued innovation in NMR data analysis. With plans to integrate the tool into routine R&D operations, AkzoNobel is well-positioned to capitalize on the efficiencies and insights enabled by data science.
Discover more about AkzoNobel by visiting their website.
You can read more of our case studies and sign up to our newsletter to keep up to date with our latest news, events and developments.