To address the challenge of optimizing laundry detergent performance, this research involved the experimental evaluation of numerous liquid detergent formulations against standardized stains. The primary aim of this project is to build a machine learning model that can accurately predict the “stain removal performance” of a detergent, given its chemical composition and other parameters, based on the performance data collected. "To remove stains" is the primary purpose of detergents, which are found in every household, and it is also the user's fundamental expectation.

For this study, 260 different detergent formulas were created. In order to cover a wide and representative chemical space pertinent to liquid detergent solutions, these formulations systematically varied in the type of surfactant, its concentration, and the amount of enzyme. Measuring cleaning effectiveness based on the agents used and stain types remains the formula developer's most effective control method, even if the physical and chemical mechanism of stain removal is fully covered in the literature. Simply explained, using a laundry detergent to remove stains involves wetting the foreign material that has been deposited on the cloth and then using the appropriate agent(s) to remove it from the surface using a chemical and/or physical process. Stain removal performance was assessed for each formulation in a controlled laboratory setting. This included certain test circumstances, such as water hardness, temperature, and wash cycle characteristics pertinent to the detergent's use, as well as typical washing machine protocols. Their measurements, stains, and tests were standardised. While conducting performance tests, we followed the A.I.S.E. detergent guidelines (4); however, because of the study's maturity, a comprehensive stain set of 19 stains was employed. It is important to note that this set of 19 stains was chosen in earlier research to create a responsive set. (5) Each formulation underwent four independent experimental repetitions to ensure reproducibility and statistical robustness of the measurements. Throughout the investigation, all stain removal performance tests were carried out using the same test parameters; the washing test protocol is provided below.

Test Procedure for Washing:

• Cotton program 40°C; 1h:49 min 1000 rpm.
• Fuzzly logic control disabled Miele W 5872 Edition 111 washing machine
• 2,5 kg Cotton and polyester ballast
• 65 mL liquid detergent
• 150 ppm CaCO₃ (15 french hardness)
• 19 stain set with 5x5 cm
• 4 repetition

The stains indicated above are standard stains produced by test materials firm like CFT and WFK. The CIELAB system was used to do a colorimetric evaluation of the stain removal performance. The colorimetric measurement's Y values are the response that was examined for the designed formula. Five raw components were employed in various proportions during the formulation design process, while ingredients that had no effect on performance were maintained constant.

Table 1. Examples of some of the tested detergent formulations. The input is represented by each column, while the formulation name is represented by each row.

By creating various detergent mixtures from scratch and testing them on a selected standartized stains, extensive data was gathered on their performance. The crucial part is to apply machine learning techniques to this data, enabling us to predict how well a new detergent will clean. Some experiments were repeated because they were deemed outliers since basic statistical markers were used during the data collection process. Testing of the machine learning model began once the dataset had reached a particular level of maturity. The goal of these tests is to determine whether the machine can "learn" the relationship or correlation between the results provided based on the given parameters. If it can, the machine will use this learning to "make predictions," which is the primary goal of this project.

The "known models" were examined as the tests were conducted and the data grew in order to assess the data and look for regression. 90% of the data was used to train the models, and the remaining 10% was utilized to assess the "prediction" capabilities in order to determine "whether the machine can learn from provided data." And the given graph reveals that “prediction” converge at earlier stages when AL(active learning) is employed.

Figure 1. Monitoring the MAE with 1-sample incremental training

AL: We start with a small subset of the available dataset and gradually grow the used set by adding examples identified by active learning along the x-axis. Performance improves as the used set grows.

Random : We start with the same small subset of the available dataset and gradually grow the used set by adding random samples along the x-axis. Performance improves as the used set grows.

According to the complexity of the dataset created, different “machine learning models” were tried in the literature and Mean Absolute Error (MAE) values were taken into account for the selection of appropriate model(s). Dataset enlargement studies continued in order to improve the prediction ability of the algorithm by selecting appropriate models.

Several "machine learning models" were tested based on the complexity of the dataset produced, and the Mean Absolute Error (MAE) values were taken into consideration while choosing the model or models. By choosing the right models, dataset enlargement studies were carried out to enhance the algorithm's prediction capability. In other words, the algorithm was requesting studies to enrich the data points that it considered “weak/insufficient”. This situation continued for a while, because the studies were done gradually, as the demands could change as new data was fed to the model. During data enrichment, improvements were made to the model structure. Two hundred and sixty different detergent formulas were designed for this study. The total number of experimental runs was 1040 (260 formulations x 4 repetitions), with 4 repetitions of the experiments performed for each of these formulations. Each of these runs yielded 19 data points to be evaluated, giving our model a rich feature set of 19760 data points.

Table 2. Performance of known models for the given data set.

The data enrichment procedure was carried out until the growth in the model improvement was deemed "insignificant", than there was a break for data enrichment. Since this model is "alive" and can be improved at any time, “break” term was used. Comparing the experimental and prediction results and determining if the predictions and experimental results are compatible are two of the project's milestones.

The algorithm was asked for the prediction results of “20 formulas that have not been tested before”, and these results were compared with the experimental results. According to this comparison, the algorithm can predict the “Y values”, but the prediction result is not the same as “numerical”. But the trend of the experimental results and the estimated results is the same. In other words, the formula with a low cleaning performance according to algorithm’s prediction will also have a bad experimental result, examples of this is shown below on a stain basis.

Figure 2. Overlaid experimental and predicted results for 20 formulations

The next stage in the project is to develop an interface where the detergent R&D team can use this algorithm. Because the aim is to have a user friendly “prediction” model and thus save both energy and time. Apart from project management, cost information sharing, and performance estimation, this portal also offered the capability of searching among the estimates based on certain criteria. Therfore, a "Formula Pool," which is not a pool made solely from estimation that individuals asked, was produced. In essence, a pool that is growing automatically is produced by the algorithm in the background depending on user-created criteria. Additionally, the price details will be provided so that a sizable pool can be accessed in the future if a search is required.

Figure 3. A screenshot of the developed interface shows how many predictions are made each month.

User generated : A member of Hayat R&D asked the model «to predict»

Auto generated : Model makes «prediction» according to Formula Pool Settings

The server went online at August 2024, as shown in the Figure 3 auto-generated “estimations” of the formulas allow user to scan formulas from larger library. To create the model, 19760 data points were gathered, as previously mentioned. Figure 3 displays the number of anticipated data points by month. Based on the user-specified parameters (ranges and increments for the five parameters), the model itself makes predictions. As mentioned before; one of the aim of this server is to scan through estimation and to find the possible formula for the given performance with intended cost. The addition of real-time pricing data for every raw ingredient in the formulation improves the predicted output even further. Through this integration, a useful database of reasonably priced and capable applicants is produced, which may be used to expedite further research and development initiatives. It would be useful to provide an example in order to highlight the "benefit" of this formula pool. One of the main problems in FMCG is unexpected (unforeseen) raw material crises brought on by cost or shortage. And it is essential to manage the issue with the least amount of harm possible without sacrificing the product's ability to remove stains, the “filter” option on the formula pool allows user for a thorough cost-benefit analysis of suggested formulations, facilitating cost and performance optimization.