INFORM April 2026

18 INFORM APRIL 2026 , VOL. 37, NO. 4

and brown grease. However, these “dirty oils” are far more challenging to process due to their variable composition, high free fatty acid levels, and need for extensive pretreatment. A recent study combined ML with PSO to optimize biodiesel production from fats, oils, and grease (FOG). FOG, collected from restaurants and food processing facilities, varies widely in composition and contains high levels of free fatty acids (nearly half), making conversion difficult. To build predictive models, researchers used a dataset of more than 20,000 experiments spanning six esterification and transesterification variables. Among eight algorithms tested, XGBoost delivered the most accurate predictions for the dataset. The team then applied PSO to further refine the model. Working within defined parameter ranges, PSO generated random input values that were fed into the XGBoost model for evaluation. This process was repeated, and after about 350 iterations, the system converged on the optimal conditions, achieving 99.4 percent conversion of free fatty acids and triglycerides to fatty acid methyl esters. Another study used ML to maximize biodiesel production from waste cooking oil using an eggshell-derived calcium oxide (CaO) catalyst. This reusable catalyst offers a cheaper, more sustainable alternative to NaOH and KOH, which require purification and generate chemical

XGBoost Predictions (R = 0.999)

100.0

97.5

95.0

92.5

90.0

87.5

Predicted Values

85.0

82.5

82.5 85.0 87.5 90.0 92.5 95.0 97.5 100.0

Actual Values

An XGBoost machine learning model showed high accuracy (R 2 =0.999) for predicting biodiesel yields under different reaction conditions. Source: Azhar, B., et al. , Energy Convers. Manag. : X, 2025.

wastewater. To prepare the catalyst, researchers washed and milled eggshells collected from restaurants, then heated them to convert calcium carbonate into CaO. The resulting eggshell-derived powder could be used without purification. The team collected 16 datasets to train and test four AI models. Waste cooking oil was first esterified with methanol and sulfuric acid to decrease free fatty acids. Then, the researchers examined how varying CaO catalyst concentration, reaction temperature, and methanol to-oil ratio affected yield during transesterification. They used 80 percent of the data for training the models and 20 percent for testing. CatBoost emerged as the top performer, identifying optimal conditions

of 3 percent catalyst, 80 °C, and a 6:1 methanol‑to‑oil ratio to produce a biodiesel yield of 95 percent. “Without AI, optimizing this green catalyst would require countless trial-and error experiments,” says Krishnamoorthy Ramalingam, postdoctoral researcher at the Universiti Sains Malysia in George Town. “Machine learning allowed us to pinpoint the most efficient reaction conditions quickly and accurately, saving time and resources.” He adds that AI models like CatBoost could be embedded in industrial control systems for real-time monitoring, fault detection, and adaptive tuning of reaction conditions. “Exploring this integration is a natural next step in our research,” he says.