A methodological framework for optimizing palm-based soap formulations using deep neural networks

This paper proposes a design framework for palm-based soap formulations that predicts key properties, melting point (MP), fatty acid profile, and cost, with high fidelity and direct applicability in industrial plants. The proposed approach enables accurate cost-performance decisions with traceability, optimizing laboratory testing, and operating within feasible windows. The approach integrates synthetic data generation under technological and process constraints, utilizing multilayer perceptron (MLP) neural metamodels for toilet soap and laundry soap. Validation is performed using correlation analysis, residual analysis, response surfaces, and local sensitivity maps. The four models (corresponding to the property and cost networks for each scenario) converge stably, reproduce experimental trends with minimal bias, maintain mean errors below 1%, and generalize well within the scope of the analyzed formulation. The MP response surfaces reveal a high plateau toward higher solid fractions (RBD palm/stearin) and a depression with increasing palm kernel olein. Sensitivity maps confirm that olein governs C12:0 (lauric acid), while palm stearin dominates C16:0/C18:0 (palmitic acid/stearic acid) and increases the MP, with olein exerting a lowering effect on this property. These findings translate into concrete lines of action in formulations, including anchoring operations on plateaus to gain tolerance to variability, applying stricter olein control in laundry soap, and leveraging flat cost gradient zones to arbitrate raw materials without compromising performance. The proposed framework optimizes experimental effort, shortens the design-validation cycle, and provides a robust basis for extending it to other lipid matrices in industrial settings.