materials-chemistry/ai in a supporting role/Food Science & Nutrition 2024 · v2
Ultrasound-treated clay bleaches cooking oil, with machine learning picking the settings
Researchers treated bentonite clay with two kinds of ultrasound and used it to strip colour from soybean and sunflower oil. Seven machine learning models were fitted to the colour measurements, and the best one guided a search for better bleaching conditions.
spectrum · one line per step, placed by what the step does · bright lines used AI
A comparative study on bath and horn ultrasound‐assisted modification of bentonite and their effects on the bleaching efficiency of soybean and sunflower oil: Machine learning as a new approach for mathematical modeling
Food Science & Nutrition, 2024
doi:10.1002/fsn3.4300 · record aix-00051 v2 · checked 2026-10-08
- AI was for
- Property prediction, Experimental design
- Model family
- Multilayer perceptron, Random forest, Support vector machine, Linear model, Gradient-boosted trees
- Checked by
- Experimental
- Code
- not reported
AI processed or interpreted data, but the main finding does not rest on it.
What this research was about
Crude vegetable oil is not the pale liquid sold in bottles. It carries natural pigments that give it a deep red or green tint, and refiners remove them in a step called bleaching. This usually means stirring the oil with a fine clay, often bentonite, which has a vast internal surface of tiny pores that pigments stick to. The more surface the clay offers, the more colour it can hold. But clay costs money, has to be disposed of afterwards, and soaks up some of the oil itself. Treating clay to make it work harder, so that less of it is needed, is a long-standing practical problem.
Here the researchers treated bentonite with ultrasound, using both a bath and a probe-like horn at different power levels, and measured the surface area and pore size of the resulting clays. They then bleached neutralised soybean and sunflower oil, varying how much clay was added, the ultrasonic power, the temperature and the time, and recorded the colour of the oil on standard scales after each run.
Where AI came in
The experiments produced a table: four settings in, five colour readings out. The researchers fitted seven machine learning regression models to this table, among them a neural network, a random forest, support vector regression, ridge and lasso regression, and two methods that build predictions from many small decision trees. All were trained from scratch on their own measurements. The task was prediction: given a set of bleaching conditions, estimate the colour of the oil that would result. The models were scored against each other and against response surface methodology, a conventional statistical way of fitting curves to experimental data. One of the tree-based methods, XGBoost, scored best on the reported measures.
That fitted model was then used as a stand-in for the experiment itself. Rather than running every possible combination of clay, power, temperature and time in the laboratory, the team let an automated search method called Bayesian optimisation probe the model, hunting within set limits for conditions that would leave the least red colour while using the least clay and energy. The model's ranking of which variables mattered most was also inspected. The settings the search returned were then run as real bleaching experiments and the measured colour compared with what the model had predicted. The reported reduction in clay use rests on this modelling step.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
Bentonite bleaching clay was treated with horn and bath ultrasound at different powers and characterised by nitrogen physisorption, and the treated clays were used to bleach neutralized soybean and sunflower oil under varying clay dose, power, temperature and time. Specific surface area rose from 31.4 ± 2.7 to 143.8 ± 3.9 m g−1 for the clay sonicated with the horn at 400 W, while mean pore diameter fell from 29.7 ± 0.14 nm to 8.3 ± 0.12 nm for the same sample. Seven multi-output machine learning regression models were fitted to the experimental colour data and compared with response surface methodology; XGBoost gave the highest R and the lowest MAE and MSE across the cases reported, with an R-test of up to .983 and an MAE of 0.498. Bayesian optimisation over the XGBoost model was used to choose bleaching conditions, and the authors report clay consumption reduced by approximately 60% for sunflower oil and 30%–35% for soybean oil.
How AI was used
Experimental bleaching runs supplied a tabular dataset whose inputs were clay percentage, ultrasonic power, time and temperature (power omitted for the control runs) and whose five outputs were Lovibond red and yellow and Hunter L*, a* and b*. Features were scaled with a min-max scaler, and seven multi-output regression models — a feedforward neural network, random forest, support vector regression, multi-task lasso, ridge regression, XGBoost and gradient boosting — were fitted to this data, with hyperparameters chosen by grid search and the non-native multi-output models wrapped in Python's MultiOutputRegressor. The fitted models predicted the colour outputs and were scored by R, MAE and MSE against a response surface methodology model built in Design Expert. XGBoost's feature importances and Pearson correlations between inputs and outputs were inspected to rank the process variables. The selected XGBoost model was then used as the objective in Bayesian optimisation, which searched within defined parameter bounds for the settings minimising the modelled red colour, clay amount and energy use; the resulting settings were run as physical bleaching experiments and the measured red colour compared with the predictions.
The shape of the work
Structural · the record, drawn
no AI
Run ultrasound-assisted bleaching experiments and characterise clay
Physical execution, by hand or by robot.
The bleaching process involved conducting experiments with activated bentonite clay added to a 50‐mL oil samplewhere the paper describes this · verbatim
no AI
Normalise input and output features
Cleaning, filtering, normalising or labelling data already obtained.
the Min–Max Scaler method is used to transform the values of the features so that they are mapped to a given rangewhere the paper describes this · verbatim
AI
Tune and fit seven multi-output regression models
Fitting model parameters, including fine-tuning an existing model. The AI stood in for statistical model.
The hyperparameters of each model were initially fine‐tuned using Grid Search to optimize their performance.where the paper describes this · verbatim
AI
Predict colour outputs and score models
Running a trained model over new data to predict, classify or score. The AI stood in for statistical model.
The best model with optimized hyperparameters was then utilized to predict the experimental data.where the paper describes this · verbatim
no AI
Examine feature importance and input-output correlations
Extracting understanding from model behaviour.
The feature importance (bleaching clay, temperature, time, and ultrasonic power) of the XGBoost regression model is presented in Figure 6.where the paper describes this · verbatim
AI
Bayesian optimisation of bleaching conditions
Iterative search over a space. The AI stood in for statistical model. Its result feeds back into an earlier step.
Bayesian Optimization efficiently explores the parameter space to identify the combination of input parameters that minimizes the model output.where the paper describes this · verbatim
no AI
Test optimum conditions experimentally
Testing outputs against ground truth.
Then the optimized samples were tested again to measure the accuracy of the XGBoost model.where the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The paper's characterisation findings (surface area, pore diameter, horn vs bath comparison) come from physical measurement; the regression models and Bayesian optimisation model and optimise the measured bleaching data, and the optimisation-derived clay saving is the one reported result that rests on the models
Seven specific regression models were selected for analysis: FNN, RF, SVR, MT Lasso, Ridge, XGBoost, and GB.where the paper describes this · verbatim
To validate the optimization results, experiments were conducted using the optimal conditions obtained from the optimization processwhere the paper describes this · verbatim
Research data are not shared.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of XGBoost (Extreme Gradient Boosting)Which version of the model was used is not stated.
- Version of Gradient BoostingWhich version of the model was used is not stated.
- Version of Random ForestWhich version of the model was used is not stated.
- Version of Feedforward Neural NetworkWhich version of the model was used is not stated.
- Version of Support Vector RegressionWhich version of the model was used is not stated.
- Version of Multi-Task LassoWhich version of the model was used is not stated.
- Version of Ridge regressionWhich version of the model was used is not stated.
About this article
Record aix-00051, version 2, checked by a person on 2026-10-08. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error