**1. Introduction**

Food production has faced important challenges in recent years, the demand for high quality foods has arisen the demands for methods to obtain rapid food manufacturing, low-level of uncertainty and error in the food formulation, safety, sustainability, and sensory satisfaction for the consumers [1]. Sensory analysis comprises a crucial step for product development because it determines in a great extent the overall acceptability, thus, the success of the product in the market. Due to the subjectivity and variability of the sensory attributes, these methods imply a large quantity of evaluations (usually >100) of trained or non-trained consumers. Therefore, a large quantity of samples is needed, which is time consuming and results in several waste of food products [2].

Artificial intelligence (AI) are computational systems, which can learn from experience and predict, classify, cluster, or even determine abstract patterns in the data provided (deep learning models). These AI methods have the advantage to manipulate a high volume of data, detect non-evident characteristics among the features, and thus, producing a projective result (prediction) in different scenarios. The application of artificial intelligence in the food area is vast, ranging from quality control (detection and determination of pesticides, proximate composition estimation, authentication of food origin, and classification of oil blends), microbiological analysis [3–6], and sensory quality (aroma profile, texture quality, among others) [1, 7].

In the sensory science field, the most used AI algorithms (models) are the partial least squares (PLS), support vector machines (SVM), random forest (RF), and deep learning approaches (artificial neural networks). These models have been successfully used for estimating sensory quality using the chemical composition [8], volatile profile of the product [9], and image analysis [10] to obtain the features for the models. However, predicting specifically the overall acceptability is challenging due to the variability among the products' formulation and characteristics.

Some regression models such as multiple linear regression (MLR), partial least squares (PLS), and regression trees have been used in the forecast of sensory quality of food products [11]. For instance, a PLS-ANN hybrid model has been reported for predicting the acceptability of green tea beverages [12], a regression trees algorithm has been applied in fresh pineapple acceptability [13] and two regression models (multiple linear regression and PLS) were used for the sensory acceptance of beef bouillon [11]. However, the models based on linear regression have some limitations such as: strong correlation among independent and dependent variables, overfitting of the model, and ignoring non-linear associations [11] which can result in a low performance (r<sup>2</sup> = 0.600–0.800).

One way to surpass the limitations of linear models is the use of artificial neural networks. Adaptive Neuro-Fuzzy Inference System (ANFIS) are known as hybrid algorithms because these models combine the benefits of neural networks (fitting and finding non-linear associations among data) and the human if-then logical learning; hence, these models can learn from data with no clear relationship among the variables and the desired output. ANFIS models have been used for predicting different type of food attributes, especially for quality control [14], drying [15], physical stability [16], rheological properties [16], but their application in sensory acceptability [17] is limited, yet it demonstrated promising results for this purpose (>90% of predictability).

Implementing AI models for estimating the sensory acceptability as a rapid screening could be useful to evaluate and reformulate food products without compromising a large quantity of samples, reducing the experimental time, or involving expensive food sensory panels. Therefore, the aim of this study was to use the formulation, the physical properties, and sensory scores to predict the acceptability of mayonnaises throughout a MLRM and an ANFIS model.
