Terry Knowles, European Correspondent04.23.25
Artificial intelligence is now among the hottest topics of interest to the coatings industry alongside sustainability, with increasing amounts of information on what AI can potentially do for the coatings sector now being offered.
This month, I offer a look at some real-time, up-to-the-minute developments around AI and machine learning and the European coatings sector.
Where AI is currently integrated into industrial systems, it is based on classical machine learning through the use of conventional computers. However, with the rise of quantum computing, which has been slowly building in the industrial background over recent years, this has the potential to create quantum machine learning, and that will be a significant leap forward in AI.
TEKNIKER, a company headquartered in Spain, has been working for more than two years in this area, with a focus on the training of AI models that will improve efficiency and accuracy in operation.
One of the major sticking points of working with conventional computing is that some problems become so big and knotted that they are almost insoluble; the quantum approach is more advanced and more ideally suited to solving more complex problems.
The researchers adopted the approach of sequential experimental design rather than the traditional design of experiment (DoE) approach; to this end, a sampling algorithm was employed iteratively to calculate and arrive at the best formulas for sample optimization, using a data-driven process based on Gaussian process modelling and Bayesian optimization.
Collectively, it combines DAST (Differential Advanced Sampling Technique), machine learning modelling, and high-throughput formulation (HTF) with a view toward improving the efficiency of coatings development.
The research was based on optimizing a 2K polyurethane coating, and the resulting data focused on the interrelationship of gloss, hiding power, scratch hardness and flexibility with regard to the complex relationships that take place between formulation parameters and intimate changes to the chemical structures incorporated through raw materials variation.
The research makes several interesting points long before reaching any conclusion and this shines a light on various factors – collective unknowns, if you like – that populate the intersection of coatings with AI or machine learning, such as:
• Many potential successful formulas remain undiscovered due to the sheer number of formulation possibilities and the inability to test them exhaustively.
• The coatings industry lags behind other chemical sectors in the application of machine learning due to the complexity of multi-component formulations and the lack of standardized data. Other areas of materials science are also in the same boat.
• The lack of public and generalized data is also a restraint; the flip side of this coin is that where data is held in proprietary databases, etc., it remains only for in-house use. Effectively, the lack of data sharing between companies hinders a much broader knowledge expansion or general consumption.
The lack of detailed data sets therefore returned the researchers to starting off with the most informative samples as a basis for model training through an approach called active learning, which reduces the need for extensive data points while improving accuracy at the same time.
Using the root mean squared error (RMSE) approach for the predicted values of coating characteristics, so that graphical interpretations of the correlations between the predicted values (x) and the actual values (y) would be perfectly diagonal, the results demonstrate a reasonable linearity for hiding power, cupping test and gloss 60° (R2 in the range of 0.86-0.88), while scratch resistance calculations and discrepancies are wider of the mark due to the complexity of the factors involved in
the formulation.
A separate analysis that examines the extent to which each type of raw material impacts each type of property showed that scratch resistance is also considerably impacted by the level of matting agent incorporation.
The conclusion of the paper supports the use of machine learning algorithms in coatings development, implying that high-throughput systems can be used not only for screening but also as an aid for modelling, effectively furnishing researchers with a deeper understanding of material characteristics.
In the context of coatings formulation, this is especially insightful when the properties of a cured coating become performance trade-offs due to the complexity of the formulation mixture. A useful advantage of this is that the coatings formulation is flexible rather than set in stone, so that formulas can be tailored according to varying
client requirements.
The new automated laboratory starts off with a base of the company’s own internal formulation knowledge and data, which will go a considerable way in optimizing contemporary coating and adhesive formulations quicker and more precisely.
A sizeable part of what will be achievable will originate from Covestro’s binders and crosslinkers that are used in coating and adhesive formulations; these are complex mixtures of chemical raw materials (about seven to 15 in total), all with varying possible outcomes and impacts on properties such as hardness, adhesion, opacity, gloss or durability, etc.
One of the objectives behind the automated laboratory will be to relieve researchers of some of the more mundane or basic aspects of coatings research, thereby freeing up research time in potentially more innovative areas, such as circularity and renewable raw materials.
The automated laboratory is expected to work around the clock and should potentially be able to test tens of thousands of samples every year. This in itself will generate huge amounts of further data that will be harnessed by Covestro, together with its already existing knowledge base to refine algorithms for better coatings and
adhesive design.
Beyond the formulation of both water- and solvent-based 1K and 2K systems and their associated formulation and film properties, application can be affected in the new laboratory under a variety of different simulated climates.
This follows in the footsteps of BYK-Chemie, which opened a high-throughput screening (HTS) facility in 2022, and the likelihood is we’ll see more of these HTE, HTS or even HTF facilities being implemented by coatings and raw materials companies alike in the future as more companies embrace machine learning in the name of efficiency, discovery and formulation/property balancing.
This month, I offer a look at some real-time, up-to-the-minute developments around AI and machine learning and the European coatings sector.
A Quantum Approach to Machine Learning from TEKNIKER
One of the key applications of AI is the optimization of industrial processes in order to achieve ideal production parameters and identification and elimination of any potential manufacturing errors or failures.Where AI is currently integrated into industrial systems, it is based on classical machine learning through the use of conventional computers. However, with the rise of quantum computing, which has been slowly building in the industrial background over recent years, this has the potential to create quantum machine learning, and that will be a significant leap forward in AI.
TEKNIKER, a company headquartered in Spain, has been working for more than two years in this area, with a focus on the training of AI models that will improve efficiency and accuracy in operation.
One of the major sticking points of working with conventional computing is that some problems become so big and knotted that they are almost insoluble; the quantum approach is more advanced and more ideally suited to solving more complex problems.
A Sequential Approach to Machine Learning in Polyurethanes
Meanwhile, in a paper currently available through Progress in Organic Coatings (https://www.sciencedirect.com/science/article/pii/S0300944025002140; Volume 205, August 2025), researchers at Niederrhein University of Applied Sciences, Institute for Coatings and Surface Chemistry (ILOC) in Germany highlight another approach to artificial intelligence for coatings testing and design.The researchers adopted the approach of sequential experimental design rather than the traditional design of experiment (DoE) approach; to this end, a sampling algorithm was employed iteratively to calculate and arrive at the best formulas for sample optimization, using a data-driven process based on Gaussian process modelling and Bayesian optimization.
Collectively, it combines DAST (Differential Advanced Sampling Technique), machine learning modelling, and high-throughput formulation (HTF) with a view toward improving the efficiency of coatings development.
The research was based on optimizing a 2K polyurethane coating, and the resulting data focused on the interrelationship of gloss, hiding power, scratch hardness and flexibility with regard to the complex relationships that take place between formulation parameters and intimate changes to the chemical structures incorporated through raw materials variation.
The research makes several interesting points long before reaching any conclusion and this shines a light on various factors – collective unknowns, if you like – that populate the intersection of coatings with AI or machine learning, such as:
• Many potential successful formulas remain undiscovered due to the sheer number of formulation possibilities and the inability to test them exhaustively.
• The coatings industry lags behind other chemical sectors in the application of machine learning due to the complexity of multi-component formulations and the lack of standardized data. Other areas of materials science are also in the same boat.
• The lack of public and generalized data is also a restraint; the flip side of this coin is that where data is held in proprietary databases, etc., it remains only for in-house use. Effectively, the lack of data sharing between companies hinders a much broader knowledge expansion or general consumption.
The lack of detailed data sets therefore returned the researchers to starting off with the most informative samples as a basis for model training through an approach called active learning, which reduces the need for extensive data points while improving accuracy at the same time.
Using the root mean squared error (RMSE) approach for the predicted values of coating characteristics, so that graphical interpretations of the correlations between the predicted values (x) and the actual values (y) would be perfectly diagonal, the results demonstrate a reasonable linearity for hiding power, cupping test and gloss 60° (R2 in the range of 0.86-0.88), while scratch resistance calculations and discrepancies are wider of the mark due to the complexity of the factors involved in
the formulation.
A separate analysis that examines the extent to which each type of raw material impacts each type of property showed that scratch resistance is also considerably impacted by the level of matting agent incorporation.
The conclusion of the paper supports the use of machine learning algorithms in coatings development, implying that high-throughput systems can be used not only for screening but also as an aid for modelling, effectively furnishing researchers with a deeper understanding of material characteristics.
In the context of coatings formulation, this is especially insightful when the properties of a cured coating become performance trade-offs due to the complexity of the formulation mixture. A useful advantage of this is that the coatings formulation is flexible rather than set in stone, so that formulas can be tailored according to varying
client requirements.
Covestro to Open Completely Automated Laboratory
Covestro has announced plans to open – sometime in 2025 – a brand-new automated laboratory that will play instrumental roles in adhesive and coatings formulation.The new automated laboratory starts off with a base of the company’s own internal formulation knowledge and data, which will go a considerable way in optimizing contemporary coating and adhesive formulations quicker and more precisely.
A sizeable part of what will be achievable will originate from Covestro’s binders and crosslinkers that are used in coating and adhesive formulations; these are complex mixtures of chemical raw materials (about seven to 15 in total), all with varying possible outcomes and impacts on properties such as hardness, adhesion, opacity, gloss or durability, etc.
One of the objectives behind the automated laboratory will be to relieve researchers of some of the more mundane or basic aspects of coatings research, thereby freeing up research time in potentially more innovative areas, such as circularity and renewable raw materials.
The automated laboratory is expected to work around the clock and should potentially be able to test tens of thousands of samples every year. This in itself will generate huge amounts of further data that will be harnessed by Covestro, together with its already existing knowledge base to refine algorithms for better coatings and
adhesive design.
Beyond the formulation of both water- and solvent-based 1K and 2K systems and their associated formulation and film properties, application can be affected in the new laboratory under a variety of different simulated climates.
This follows in the footsteps of BYK-Chemie, which opened a high-throughput screening (HTS) facility in 2022, and the likelihood is we’ll see more of these HTE, HTS or even HTF facilities being implemented by coatings and raw materials companies alike in the future as more companies embrace machine learning in the name of efficiency, discovery and formulation/property balancing.
