Inking In The Future for Machine Learning in Formulation Design

Inks, paints, coatings – these may not be the first things that you think of when you hear ‘artificial intelligence’ or ‘machine learning’. But optimising these chemical formulations is a problem of high commercial significance that also pushes the boundaries of data analysis.  Printing technology leader Domino Printing Sciences teamed up with machine learning specialist, Intellegens, to apply novel ‘deep learning’ techniques, dramatically reducing experimental timescales.

Ink properties like colour, flow, adherence to surfaces, and response to light depend on subtle changes in the nature, proportions, and processing of ingredients. The challenge of understanding these subtleties is similar for other formulated products and plastics. Manufacturers want to control properties not only to make new products, but also to re-formulate in response to market pressures – for example, replacing an ingredient that has become expensive or that is subject to new environmental regulations.

The traditional approach is time-consuming and expensive: a systematic experimental program that tests variations in the formulation, identifying those that produce the best results. Data from past experiments is used to design these programs, identifying relationships between properties and formulation features to suggest what changes should be tried. But the available data is usually sparse. That is, if you imagine it in a spreadsheet, then few, if any, of its rows and columns are complete. It was gathered over many different projects and formulations, with each individual experiment focused on immediate objectives. It usually doesn’t make sense to measure every possible property for every sample.

This limits what conventional analysis tools – even machine learning methods – can extract from the data. They need a critical mass of complete ‘rows’ or ‘columns’. This limitation is amplified when trying to optimize many different parameters. Typical machine learning algorithms learn from inputs and outputs. When too many inputs are missing, they fail. Most of these approaches also require significant data cleaning. Intellegens ‘deep learning’ technology, Alchemite™, was developed by University of Cambridge physicists to overcome these restrictions. Its advanced techniques learn from sparse, noisy data while solving multi-parameter problems. It requires minimal data cleaning. Alchemite™ makes sensible predictions for gaps in the data and calculates the uncertainty associated with each prediction. It recommends which experiment to do next, and can then use the new data to continuously improve its model.

Domino uses Alchemite to optimise ink formulations, decrease time-to-market, and release valuable lab resources.  Avoiding unnecessary tests has enabled them to reduce some experimental timescales from months to minutes. Dr Andrew Clifton, Director of Marking Materials and Test Engineering, commented “We were impressed with the ability of Alchemite™ to identify novel formulations quickly and accurately. This enabled us to make the most of limited lab resources and continue innovating during the COVID-19 lockdown.”

Read the full case study and hear Domino in a recorded webinar at