1. Bibliographic Information

1.1. Title

The title of the paper is A step forward in food science, technology and industry using artificial intelligence. It clearly indicates the central topic: the application and advancement of artificial intelligence (AI) in various aspects of the food sector.

1.2. Authors

The authors are Rezvan Esmaeily, Mohammad Amin Razavi, and Seyed Hadi Razavi. Their affiliations are indicated with an address associated with "Box 4111, Karaj, 3158777871, Iran", suggesting a research institution or university in Iran. The presence of multiple authors, including a corresponding author (Seyed Hadi Razavi, denoted by *), is typical for academic research, indicating collaborative work.

1.3. Journal/Conference

The paper was published in a journal, as indicated by the presence of a "Handling Editor: Dr. S Charlebois" and the standard journal article structure. Although the specific journal name is not explicitly stated in the provided text snippets beyond the abstract and introduction, the reference format (e.g., https://doi.org/10.1016/j.tifs.2023.104286 in the Data availability section and the References section) strongly suggests it was published in Trends in Food Science & Technology. This is a highly reputable and influential journal in the field of food science, known for publishing high-impact review articles and original research that shapes the discourse in food technology and related industries.

1.4. Publication Year

The Digital Object Identifier (DOI) $10.1016/j.tifs.2023.104286$ and several references within the text (e.g., $(2023)$) indicate that the paper was published in 2023.

1.5. Abstract

The paper's background emphasizes the vital role of food science, technology, industry, processing, nutrition, functional foods, and nutraceuticals for human health. It highlights the importance of agriculture as the primary food source, necessitating the removal of hazardous chemicals and microorganisms, ensuring food quality and safety through subsequent processing, and evaluating food intake for nutrition. The scope and approach involve integrating advanced technologies, particularly artificial intelligence (AI) and machine learning (ML), into academic and industrial food domains. AI applications are shown to improve agriculture (e.g., crop classification), aid in formulation, develop new food products, enhance sensory evaluations (using electronic nose, electronic tongue, and machine vision), improve industrial processing, assure food quality and safety, manage supply chains, facilitate waste reuse, and estimate calorie and nutrient content. The key findings and conclusions assert that deploying AI algorithms across agriculture, food science, and nutrition can establish an intelligent from-farm-to-fork system, significantly advancing the scientific field.

1.6. Original Source Link

The original source link provided is /files/papers/6908b649e81fdddf1c48bfc2/paper.pdf. This appears to be a local file path or an internal identifier within a larger system, indicating that the content is derived from a specific PDF document. Based on the DOI $10.1016/j.tifs.2023.104286$ mentioned in the Data availability section, the paper is officially published in Trends in Food Science & Technology.

2. Executive Summary

2.1. Background & Motivation

The core problem the paper addresses is the need for continuous advancement in the food sector to ensure a healthy lifestyle, food quality, and safety from farm-to-fork. Food is fundamental for human life, making food science critically important globally. This field encompasses diverse areas like food technology, industrial processing, nutrition, functional foods, and nutraceuticals. Agriculture, as the primary source of food, faces challenges such as the presence of hazardous chemicals and microorganisms. Subsequent food processing steps must guarantee safety and quality, while evaluating food intake is crucial for nutrition.

This problem is important due to global challenges such as population growth, climate change, and the increasing demand for high-quality, safe, and sustainable food. Traditional methods for addressing these issues are often time-consuming, less accurate, and unable to handle the vast amounts of data generated throughout the food supply chain.

The paper's entry point is the increasing popularity and capability of advanced technologies, particularly artificial intelligence (AI) and machine learning (ML), in both academic and industrial domains. It posits that these technologies offer solutions to the complex, data-intensive problems within food science and technology, presenting an opportunity to revolutionize how food is produced, processed, and consumed. The innovative idea is to systematically integrate AI throughout the entire from-farm-to-fork continuum, creating an intelligent, efficient, and safer food system.

2.2. Main Contributions / Findings

The paper's primary contributions lie in its comprehensive synthesis and articulation of how AI and ML can be applied across the entire food value chain. It doesn't propose a new model or algorithm but rather provides a structured overview of existing and emerging AI applications, thereby defining a roadmap for the intelligent transformation of the food sector.

The key conclusions and findings are:

• Holistic From-Farm-to-Fork Integration: AI algorithms can be employed throughout agriculture, food science, and nutrition to create a fully intelligent from-farm-to-fork system. This integrated approach is a significant advancement for the scientific field.

• Diverse Application Areas: AI and ML applications are instrumental in:

  • Agriculture: Improving farming practices, crop classification, and managing hazardous chemicals and microorganisms.
  • Product Development: Facilitating formulation, developing new food and nutraceutical products.
  • Sensory Evaluation: Enabling more accurate assessments through electronic nose, electronic tongue, and machine vision.
  • Industrial Processing: Enhancing efficiency and control in food processing.
  • Quality & Safety Assurance: Assuring food quality and safety throughout the chain.
  • Supply Chain Management: Optimizing supply chains for efficiency and reduced waste.
  • Sustainability: Promoting waste reuse.
  • Nutrition: Estimating calorie and nutrient contents and supporting personalized nutrition.

• Overcoming Challenges: AI offers advantages such as low-cost requirements, user-friendliness, improved performance, accuracy, speed, powerful analysis, modernity, eco-friendliness, and time-saving capabilities, which generally outweigh its potential drawbacks.

• Future Potential: The nascent stage of AI integration in many sub-domains of food science indicates vast potential for numerous research avenues and breakthroughs, especially given its relevance to basic human needs.

  These findings collectively address the problem of inefficiency, safety concerns, and quality control in the food sector by proposing a data-driven, intelligent framework powered by AI.

3. Prerequisite Knowledge & Related Work

3.1. Foundational Concepts

To understand this paper's comprehensive review, several foundational concepts related to artificial intelligence, machine learning, and sensors are crucial.

• Artificial Intelligence (AI): Artificial intelligence (AI) is a broad field of computer science that aims to create intelligent machines capable of performing tasks that typically require human intelligence. This includes learning, problem-solving, decision-making, perception (e.g., visual, auditory), and language understanding. The paper notes that AI systems are designed to model how humans solve problems and translate that logic into computer language. It also highlights that AI can simulate or even improve upon parts of human brain function. The term was coined by Dr. John McCarthy in 1956. Key applications within AI include pattern recognition (PR), voice detection (VD), digital signal processing (DSP), machine learning (ML), and natural language processing (NLP). AI draws upon diverse fields such as computer science, mathematics, statistics, robotics, and sensor-related sciences.

• Machine Learning (ML): Machine learning (ML) is a core subset of AI that focuses on enabling systems to learn from data without being explicitly programmed. Instead of hard-coding rules, ML algorithms use statistical methods to identify patterns and relationships within data, allowing them to make predictions or decisions. The paper categorizes ML into four types:

  1. Supervised Learning: This type of ML uses labeled training data, meaning the input data is paired with the correct output. The algorithm learns to map inputs to outputs and then generalizes this knowledge to new, unseen data. Examples include neural networks (NN), logistic regression, support vector machines (SVM), K-nearest neighbor (KNN), and random forest (RF). It's further divided into classification (predicting a categorical label) and regression (predicting a continuous value).
  2. Unsupervised Learning: Unlike supervised learning, unsupervised learning works with unlabeled data. Its goal is to discover hidden patterns, structures, or groupings within the data. Common methods include clustering (grouping similar data points, e.g., K-means) and dimension reduction (reducing the number of variables while retaining important information, e.g., Principal Component Analysis (PCA)).
  3. Semi-supervised Learning: This approach combines aspects of both supervised and unsupervised learning, using a small amount of labeled data along with a large amount of unlabeled data for training.
  4. Reinforcement Learning (RL): In RL, an agent learns to make decisions by performing actions in an environment and receiving rewards or penalties based on those actions. The goal is to maximize the cumulative reward over time through trial and error. It often follows Markov-chain rules and can achieve customized algorithms. Examples include generative adversarial networks (GAN) and transformers.

• Deep Learning (DL): Deep learning (DL) is a specialized subset of ML that utilizes artificial neural networks (ANN) with multiple layers (typically three or more). These deep networks are capable of learning complex hierarchical representations of data, making them particularly effective for tasks like image recognition, natural language processing, and handling very large datasets. The paper mentions convolutional neural networks (CNN) as a common type of deep neural network, especially for image processing. A key characteristic of DL is its ability to automatically extract features from raw data, unlike some traditional ML methods where features need to be manually engineered.

• Sensors: Sensors are devices that detect and respond to changes in their environment. They convert physical stimuli (e.g., temperature, pressure, light, chemical presence) into electrical signals that can be read and processed by electronic systems. In the context of food science, specialized sensors play a critical role in quality control, safety assurance, and sensory evaluation. The paper highlights several types:

  • Direct Sensors: These use materials like metal, metal oxide, metal-organic frameworks (MOFs), solid electrolytes, carbon nanotubes, graphene, nano-materials, and nano-particles without chemical reactions or energy format transformation.
  • Smart Sensors: These are advanced sensors, including nuclear sensors (for mass flow, medical imaging), micro sensors (various materials for medical/automotive), and nano-sensors (for biological/chemical uses, due to size-dependent properties). MEMS-based sensors (micro-electro-mechanical systems) are an example of microsensors.
  • Passive Sensors: Classifiable by parameters they recognize, such as position, temperature, force, flow, and pressure.
  • Complex Sensors: Combinations of multiple sensors for multi-parameter detection (e.g., temperature sensors combined with other passive sensors).
  • Electronic Nose (E-nose) and Electronic Tongue (E-tongue): These are AI-based devices composed of arrays of gas or chemical sensors, respectively, designed to mimic human olfaction and gustation for rapid and accurate detection of aromas and tastes. They are often integrated with ML algorithms like SVM, PCA, and NN for data analysis.
  • Machine Vision: A technology that enables computers to "see" and interpret images, often used for quality inspection and classification in food.

• Internet of Things (IoT) & Internet of Food (IoF): The Internet of Things (IoT) refers to a network of interconnected physical devices, vehicles, home appliances, and other items embedded with sensors, software, and other technologies that allow them to connect and exchange data over the internet. In the paper, it emphasizes continuous data transfer between diverse devices, software, and sensors, typically via a wireless sensor network, without public internet access restrictions. The Internet of Food (IoF) is a specific application of IoT within the food sector, where relevant data and models are shared across a network to improve quality systems at different stages of food production, aiming for greater efficiency and sustainability.

3.2. Previous Works

The paper, being a review, synthesizes numerous previous works rather than presenting a single new one. It frames the evolution of AI applications from historical roots to modern use cases in the food sector.

• Early AI Development (1950s): AI research dates back to 1950 or earlier, initially used for simple problems in mathematics and computer science. Later, it developed to recognize patterns.

• Resurgence with Big Data (Recent Decades): The widespread availability of big data and advancements in algorithms led to a resurgence of AI applications. The popularity of machine learning in industry grew significantly due to progress in ML and the increasing size of industrial data.

• Broad Adoption Across Disciplines: The necessity of AI and ML is growing continuously across all fields, including chemistry and crystallization. A combination of AI and biotechnology is seen as crucial for sustainable development in areas like climate, food, health, and natural resources.

• Specific AI Applications cited in the Introduction:

  • Bioprocess Engineering: AI offers potential to improve bioprocess engineering through process and design optimization.
  • Food Processing Cycle: AI plays a significant role in the modern food processing cycle, from sorting to analysis, packaging, customer satisfaction, and supply chain management.
  • Handling Large Data: AI is essential for assaying large-sized data and complex variables in food production, from lab to industrial scale, while also offering time-saving benefits.
  • IoT, Sensing, ML, Data Analytics: These technologies have driven improvements in food engineering, production, quality, safety, food science discovery, waste management, and supply/demand chain management.
  • Agriculture: Farmers are using analytical tools based on AI, neural networks for crop prediction, and ML/deep learning to select and classify agri-products based on soil type. Smart agriculture uses AI to improve issues related to crops, soil, water, and other agricultural processes. AI models are used to predict food risks related to climate change.
  • Medical and Nutrition Fields: AI is leading in prevention, diagnosis, medical imaging, treatment methods, and human nutrition. AI is used for food classification and micronutrient analysis. IoT and AI are expected to bring tremendous progress in functional food.
  • Personalized Diet: AI, particularly ML methods, is effective for detailed food item analysis and customized diets. An example includes AI algorithms for tracking meals of type 1 diabetic patients to estimate carbohydrate percentage.

• Foundational Formulas in Related Work: Given that this paper is a review that synthesizes existing applications of AI/ML rather than introducing a novel AI/ML method or formula itself, it primarily describes the concepts and applications of these foundational AI/ML techniques. Therefore, the paper does not present specific formulas for Attention mechanisms or other complex ML models within its Related Work section (or its Methodology section, which focuses on describing these techniques conceptually). The expectation for prerequisite knowledge is to define these terms, not to derive their mathematical underpinnings from external sources unless the current paper specifically builds upon those formulas in a novel way. The core algorithms are explained conceptually within the paper, and I will detail those explanations.

3.3. Technological Evolution

The technological evolution leading to the current state of AI in food science can be traced through several phases:

1. Early AI (1950s): AI began with efforts to solve simple mathematical and computer science problems. This phase laid the groundwork for logical reasoning and computational capabilities.

2. Pattern Recognition (Later Years): The application of AI evolved to address pattern recognition, tackling problems that are visually easy for humans but difficult to describe algorithmically. This was a crucial step towards tasks like image processing and data classification.

3. Big Data Era and ML Advancement: The exponential growth of data (big data) combined with significant advancements in machine learning algorithms, particularly deep learning, marked a turning point. This allowed AI to handle large datasets and complex variables more effectively, leading to its widespread adoption. The paper notes the "huge progress in machine learning as well as growing the size of industrial data" as the reason for ML's popularity.

4. Integration of AI with IoT and Sensors: The development of Internet of Things (IoT) and sophisticated sensors (e.g., E-nose, E-tongue, machine vision) created a rich ecosystem for AI. IoT enables continuous data transfer, while sensors provide the real-time data needed for AI models to monitor and control processes in agriculture and food production.

5. From Specific Tools to Holistic Systems: Initially, AI applications might have been siloed (e.g., neural networks for crop prediction). The current evolution, as highlighted by the paper, moves towards integrating AI into an intelligent from-farm-to-fork system, where AI manages every stage from agriculture to consumption, encompassing quality, safety, supply chain, waste management, and nutrition. This represents a shift from solving isolated problems to building comprehensive, interconnected intelligent systems.

   This paper's work fits within the latter part of this timeline, focusing on the current state and future potential of AI as a holistic enabler for the entire food ecosystem.

3.4. Differentiation Analysis

As a review paper, this work does not propose a new methodology to differentiate from existing ones. Instead, its core innovation lies in its comprehensive scope and systematic integration perspective.

Compared to other research that might focus on specific AI applications (e.g., ML for crop yield prediction or E-nose for food quality), this paper differentiates itself by:

• Holistic From-Farm-to-Fork Framework: The primary differentiation is the emphasis on creating an "intelligent from-farm-to-fork system." While individual AI applications in food science exist, this paper synthesizes them into a continuous, interconnected workflow spanning agriculture, processing, safety, distribution, and nutrition. This broader vision provides a structured overview that is valuable for both researchers and industry practitioners looking to implement AI strategically.

• Cross-Domain Synthesis: The paper brings together AI applications from diverse and often separately studied domains—agriculture, food science, food industry, nutraceuticals, and human nutrition—into a single coherent narrative. This interdisciplinary approach highlights the synergistic potential of AI across these fields.

• Identification of Opportunities and Challenges: Beyond merely listing applications, the paper also discusses the opportunities and challenges associated with AI implementation, providing a balanced perspective that is crucial for practical adoption.

• Emphasis on Enablers: It dedicates sections to foundational AI concepts (ML, DL), IoT, and sensors, positioning them as key enablers for the described AI advancements in food, thereby providing a more complete picture of the technological landscape.

• Beginner-Friendly Description of Core AI/ML Methods: For a review, it provides concise yet clear explanations of the core ML algorithms (PCA, SVM, DT, RF, NN, DL, RL), making the paper accessible to readers who might be experts in food science but novices in AI/ML.

  In essence, the paper's innovation is its role as an integrator and visionary, mapping out how existing and evolving AI technologies can collectively elevate the entire food sector, rather than introducing a single novel technical solution.

4. Methodology

The "methodology" of this paper is its structured approach to reviewing and synthesizing the applications of artificial intelligence (AI) and machine learning (ML) in food science, technology, and industry. It systematically breaks down AI concepts, ML algorithms, enabling technologies like sensors and IoT, and then details their applications across the food value chain. The paper does not propose a novel AI algorithm or experimental methodology but rather serves as a comprehensive overview of existing techniques and their deployment.

4.1. Principles

The core idea driving this review is that AI can significantly enhance the food cycle from farm-to-fork. The theoretical basis is that by modeling human intelligence processes and leveraging computational power, AI can analyze complex data, recognize intricate patterns, make predictions, and automate tasks more efficiently and accurately than traditional methods. This efficiency translates into improvements in food quality, safety, sustainability, and economic viability. The intuition is that just as AI has revolutionized other sectors, its systematic application in the data-rich and complex food domain can lead to a paradigm shift, creating intelligent systems that optimize every stage of food production and consumption.

4.2. Core Methodology In-depth (Layer by Layer)

The paper's methodological review structure can be broken down into the following components:

4.2.1. Introduction to Artificial Intelligence (AI)

The paper begins by defining Artificial Intelligence (AI) as the field where state-of-the-art computer systems perform duties traditionally executed by natural (human) intelligence. These systems model the human brain and make analysis to perform reasoning tasks. The process involves reviewing how humans solve a problem and translating that logic into computer language using codes.

AI encompasses several key topics:

• Pattern Recognition (PR)

• Voice Detection (VD)

• Digital Signal Processing (DSP)

• Machine Learning (ML)

• Natural Language Processing (NLP)

  The foundation of AI is interdisciplinary, drawing from computer science, mathematics, statistics, robotics, and sensor-related sciences.

4.2.2. Machine Learning (ML)

Machine Learning (ML) is presented as a subset of AI that empowers machines to automatically learn from given data by implementing statistical algorithms and models. ML algorithms are designed to train machines and systems to automatically identify relationships and patterns and make predictions.

ML is categorized into four main types:

4.2.2.1. Supervised Learning

• Principle: Uses a training dataset with labeled data to teach the machine. Labeled data means that the relations between features and desired output are marked.
• Process: Algorithms iteratively learn over time and optimize its performances during each step.
• Sub-types: Classification (predicting categories) and regression (predicting continuous values).
• Common Algorithms: Neural networks (NN), logistic regression, support vector machines (SVM), K-nearest neighbor (KNN), and random forest (RF).

4.2.2.2. Unsupervised Learning

• Principle: Works on unlabeled data without known outputs.
• Purpose: To discover hidden patterns, or perform grouping and clustering based on similarities or differences.
• Common Methods: Clustering (K-mean) and dimension reduction (PCA).

4.2.2.3. Semi-supervised Learning

• Principle: Not explicitly detailed in the paper's main text for this section, but listed as a category, implying a combination of labeled and unlabeled data.

4.2.2.4. Reinforcement Learning (RL)

• Principle: Utilized for non-convex math functions that lack a unique single optimum point.
• Mechanism: Follows Markov-chain rules. It doesn't have a specific predefined algorithm but achieves its customized algorithm by doing trial and error several times.
• Components (RL Agents): Actor, state, environment, and action. An actor transitions from one state to another, receiving a reward. The amount and level of the reward are variable based on accuracy, driving the machine/robot to maximize reward.
• Forms: Generative adversarial network (GAN) and transformers.

4.2.3. Deep Learning (DL)

Deep Learning (DL) is defined as a subset of ML that employs artificial neural networks (ANN) with complex multilayers (more than two layers). Convolutional Neural Network (CNN) is given as an example of a common deep neural network. A key distinction is that in DL, features are automatically entered for more complex algorithms, unlike some AI methods requiring manual feature entry. DL requires large datasets due to its advanced computing system.

4.2.4. Specific ML Methods and Algorithms (Section 2.1.1 in paper)

4.2.4.1. Principal Component Analyses (PCA)

• Function: The most popular dimension reduction algorithm.
• Type: A subset of unsupervised learning.
• Mechanism: Transforms a large set of variables into smaller dimensions while keeping variance maximum and without losing important information.
• Application: Commonly used in E-nose.

4.2.4.2. Support Vector Machine (SVM)

• Applications: Wide range, including chemical classification, data mining, bioinformatics, and face recognition.
• Mechanism: Relies on statistical learning theory (SLT). It uses a kernel function (e.g., polynomial, sigmoid, Gaussian kernel, radial basis function (RBF), linear, and nonlinear) to transfer data from the input environment to a higher-dimensional space.

4.2.4.3. Decision Tree and Random Forest (DT and RF)

• Decision Tree (DT): A classifier characterized by external and internal nodes arranged in several layers and connected by branches. Internal nodes contain a decision function, while external nodes indicate the output for an input vector. The ID3 algorithm is a popular DT variant.
• Random Forest (RF): Formed when various trained decision trees collect to each other to achieve a single outcome. Useful for both classification and regression issues.

4.2.4.4. Neural Network (NN) / Artificial Neural Network (ANN)

• Inspiration: Computational method inspired by the biological nervous system function of the human brain.
• Capabilities: Handles complex problems in diverse fields (e.g., agriculture, science, medical science, finance, engineering, art). Can be used for image recognition and natural language processing.
• Advantages: Reduces calculation complexity and cost, handles larger input volumes.
• Structure: Neural network layers are independent. Each layer contains processing nodes (neurons) linked by weighted connections. A bias is a hyperparameter that helps the network adjust activation functions. Weights and biases are adjusted during training for data classification.
• Applications: Pattern recognition, prediction, classifications. Applicable in all ML approaches (supervised, unsupervised, semi-supervised, reinforcement learning).
• Considerations: NN performance is dependent on time and hardware, necessitating appropriate methods like Convolutional Neural Network (CNN).
• Convolutional Neural Network (CNN): A type of deep learning neural network primarily applied in image processing. Output is generated by processing image data through a group of convolutional layers.

4.2.4.5. Deep Learning (DL)

• Definition: Reemphasized as a subset of ML, referring to artificial neural networks (ANN) with complex multilayers (more than two layers). CNN is a common example.
• Feature Handling: DL automatically inputs features into complex algorithms, unlike some AI methods requiring manual feature entry.
• Data Requirements: Requires large datasets due to its advanced computing system.
• Recent Trend: The past decade has seen a peak in deep-learning applications for object identification.

4.2.4.6. Optimization of Deep Learning

• Mathematical Context: Concerns optimal points in functions.
  • Convex functions: Have a single optimum point. Machine learning (ML) is suitable.
  • Non-convex functions: Have some optimum points. Reinforcement learning (RL) is generally used.
• ML Optimization Process (for convex functions):
  1. ML trains in shuffle order and iterates for myriad times.
  2. Cost function is measured (using methods like RMSE, MSE, MA).
  3. Error percentage is decreased through optimization using algorithms such as SGD (stochastic gradient descent), Adam, and back propagation.
  4. Cost function is remeasured.
• Metrics for Cost Function:
  • RMSE (Root Mean Square Error)
  • MSE (Mean Squared Error)
  • MA (Mean Absolute Error)

4.2.4.7. Reinforcement Learning (RL)

• Application: Used for non-convex math functions that do not have a unique single optimum point.
• Core Principle: Based on Markov-chain rules. RL can derive its customized algorithm through trial and error.
• Types of RL: Generative Adversarial Network (GAN) and transformers.
• Agent Components: Actor, state, environment, action. An actor moves from state 0 to state 1 and receives a reward. This action is iterated across states. The reward amount varies based on the accuracy of diagnosis and task performance, motivating the machine/robot to maximize rewards.

4.2.5. Internet of Things (IoT) and Internet of Food (IoF)

• IoT: Involves the continuous transferring of data among diverse devices, software, sensors, and technologies. Requires a wireless sensor network. It operates without restrictions, enabling data delivery to equipment or individuals anytime and anywhere, linking them through a smart and customized network without necessarily accessing the public internet. IoT is growing in areas like smart agriculture, smart healthcare, smart home, and smart cities.
• IoF: A specialized application where relevant data and models are shared in a network to improve quality systems at different stages of food production. Aims for greater efficiency and sustainability in the food industry.

4.2.6. Sensors

The paper discusses sensors based on their materials and types, which are crucial for data collection in AI-driven food systems. The classification is summarized in Figure 2.

该图像是一个传感器分类示意图，展示了被动传感器、复杂传感器以及智能传感器的分类和关联，重点突出了材料（如纳米材料）对传感器的影响及其在食品科学中的应用。

Figure 2: Classification of all types of sensors. Abbreviation: refer to appendix.

• Materials for Direct Sensors:

  • Metal: e.g., gold in biosensors, platinum in automobile engines and medicine.
  • Metal oxide: Used as biosensors (thin film or thick compressed powder).
  • Metal-Organic Frameworks (MOFs): Crystalline, less-density structures, used for separation and filtration of gases.
  • Solid electrolytes: e.g., oxygen sensors.
  • Graphene and carbon nanotubes: Form electrochemical, chemical, acoustic, environmental, resonance, optical, humidity, gas, and mechanical sensors.
  • Nanomaterials and nano-particles: Distinguished by their size ($\leq 100 \ \mathrm{nm}$), leading to changed physicochemical properties. Nano-sensors are widely used in chemical and biological fields.

• Biosensors: Material choice significantly impacts characteristics. AI biosensors are critical for disease treatment, leveraging AI algorithms such as SVM, PCA, LDA, and ANN, often implemented with E-nose. Protein and peptide-derived biosensors (from animal olfactory systems) are an advanced type for artificial olfaction (E-nose).

• New Generation (Smart Sensors):

  • Nuclear sensors: Based on Gamma rays for mass flow and medical imaging.
  • Micro sensors: Made from various materials (gases, metal, plastic, polymers, ceramics) for medical devices and automotive industry. MEMS-based sensors are a subtype, notable for electrical/mechanical properties and scale.
  • Nano-sensors: Using metal film and semiconductors (thin layer form) for biological and chemical uses. Fluorescent nano-sensors (especially carbon-containing) are used for monitoring food quality, detecting compounds like functional chemicals, heavy metals, pesticide residues, veterinary drugs, illegal additives, toxins, and microorganisms. Carbon dot nano sensors offer superiorities over traditional fluorescent ones.

• Passive Sensors: Classified by parameters they recognize: position, temperature, force, flow, and pressure.

• Complex Sensors: Combinations of multiple sensors for multi-parameter detection. Examples include temperature sensors combined in thermistor, thermocouple, and resistance thermometer. Others exist for light, humidity, motion, and mass sensing.

4.2.7. Role of Dataset Characteristics

The paper emphasizes that significant factors such as size, type (binary, categorical, or continuous), and labeling (labeled or unlabeled) of datasets are crucial in choosing the appropriate algorithm and classifier.

The paper provides a table (Table 1) summarizing specifications of supervised algorithms for non-linear applications.

The following are the results from Table 1 of the original paper:

This table offers a concise comparison of ANN, SVM, and DT based on various performance and dataset characteristics, guiding the selection of algorithms based on specific application needs.

5. Experimental Setup

This paper is a comprehensive review article rather than a primary research paper presenting new experimental results. Therefore, it does not describe an experimental setup in the traditional sense, with specific datasets, evaluation metrics, or baselines designed and implemented by the authors themselves. The authors' methodology was to synthesize existing research.

However, the paper does discuss how AI and ML methods are applied in various food science, technology, and industry (FSTI) contexts, drawing from numerous studies conducted by other researchers. These applications implicitly involve specific experimental setups, which the paper summarizes conceptually.

5.1. Datasets

The paper does not use specific datasets in its own research. Instead, it refers to the types of data used in the applications it reviews:

• Chemometric Data: Data derived from analytical methods such as chromatography, electrophoresis, spectroscopic methods, and hyperspectral imaging (HSI). These are often subjected to unsupervised analysis.

• Sensory Evaluation Data: Data from odor and flavor analysis, which can be acquired via E-nose and E-tongue, often involving chemical or gas sensor arrays.

• Agricultural Data: Data related to crops, soil, water, climate conditions, pesticide residues, and hazardous chemicals.

• Food Product Data: Data concerning quality characteristics (e.g., transparency, color, foam, stability, body for beer; ripeness for fruits), adulteration, contamination, freshness, nutrition content, and waste materials.

• Image Data: For machine vision applications (e.g., CIP systems, categorizing fresh products, saffron adulteration, cocoa beans, walnut trees), hyperspectral images, and smartphone-uploaded images.

• Consumer Data: For consumer analytics (e.g., demand, perception, purchasing behavior).

• Time-series Data: For process control, shelf-life prediction, and climate change impacts.

• Genetic and Environmental Data: For livestock health and breeding.

  The paper highlights that AI and ML are particularly useful for assaying large-sized data and complex variables that are common in food production. While it doesn't provide concrete examples of data samples from a specific experiment, it implies the use of diverse multimodal data (chemical, physical, visual, sensory, environmental) across the different application areas.

5.2. Evaluation Metrics

The paper describes the applications of AI/ML but does not define specific evaluation metrics for its own review process. However, it implicitly refers to the goals of evaluation in the reviewed studies. For example, in the Optimization of Deep Learning section, it mentions common metrics used to measure the cost function and decrease error percentage during ML training:

• RMSE (Root Mean Square Error)

• MSE (Mean Squared Error)

• MA (Mean Absolute Error)

  When discussing the performance of algorithms, it mentions:

• Accuracy: This is a common metric in classification tasks, representing the proportion of correctly classified instances out of the total instances.

  • Conceptual Definition: Accuracy measures the overall correctness of a model's predictions. It indicates how often the model is correct across all classes.
  • Mathematical Formula: $ \text{Accuracy} = \frac{\text{Number of Correct Predictions}}{\text{Total Number of Predictions}} $
  • Symbol Explanation:

    • Number of Correct Predictions: The count of instances where the model's prediction matches the true label.

    • Total Number of Predictions: The total number of instances evaluated by the model.

      In Table 1, Overall Accuracy is listed as a Performance Item for ANN, SVM, and DT. For LDA in the beer classification example, it states LDA had a more accurate performance than PCA, again implying the use of an accuracy metric.

5.3. Baselines

As a review, the paper does not establish its own baselines. Instead, it synthesizes findings where other researchers have compared AI/ML methods against traditional methods or other AI/ML algorithms. Examples mentioned include:

• Traditional Methods: AI-based technologies are seen as improvements over traditional methods (e.g., for sensory evaluation, food quality detection). The paper notes that analytical methods like chromatography, electrophoresis, spectroscopic methods, and hyperspectral imaging (HSI) have replaced sensory assessments to prepare authentic foods.

• Other AI/ML Algorithms: Comparisons are highlighted between different ML algorithms, such as LDA versus PCA for beer classification or FFNN vs. SVM, LDA, LR, iPLS for sugar concentration estimation. In some cases, DL is shown to be more successful than classical methods for wheat variety recognition.

  The intent is to showcase the superior performance or specific advantages of AI/ML in various contexts, implicitly using non-AI or less advanced AI methods as comparison points.

6. Results & Analysis

The paper's results and analysis are presented as a comprehensive overview of AI and ML applications across the food science, technology, and industry (FSTI) domain, rather than experimental outcomes from the authors' own research. The key insights are derived from synthesizing findings from numerous studies.

6.1. Core Results Analysis

The paper structures its findings by AI's impact on different aspects of FSTI:

• AI in FSTI (General Applications): AI plays an impressive role in food science and industry, including making CIP (cleaning in place) systems efficient, managing supply chains, developing new products based on consumer preferences, categorizing fresh products, ensuring food quality, controlling processes, processing images, evaluating sensory aspects (odor, flavor), and saving time and resources. Crucially, AI is extended for prediction in food safety and foodborne diseases. The benefits of AI—low-cost, user-friendly, improved performance, accuracy, speed, powerful analysis, modernity, eco-friendliness, time-saving—are stated to outweigh its potential drawbacks. Food fingerprints (biomarkers for product characteristics) are effectively analyzed by AI methods.
• ML in FSTI and Sensory Evaluation: ML is confirmed for modeling and data analysis in food science. It's used for food characterization analysis, particularly chemical cases, and to a lesser extent, sensory and texture aspects. Variables from chemical, agricultural, physical, nutritional, and sensory evaluation parameters are analyzed by AI-based technologies like E-nose and E-tongue, or professional devices. E-nose and E-tongue, devices mimicking human senses, provide accurate, inexpensive, and instant food quality estimation using SVM, PCA, PLS, and neural networks. They can detect five main flavors, trillion odors, and distinguish adulterants and toxic chemicals. Machine vision also contributes to sensory evaluation.
• AI and Food Formulation: ML (e.g., non-negative matrix factorization, two-step regularized least squares) is increasingly used for formulation, allowing for recipe integration. ML can also enrich products, make them healthier, or remove/improve items for consumers with intolerances/allergies, leading to specialized outcomes. The performance is highly dependent on data quality.
• AI and Food Waste: AI aids in the reuse of waste materials, improving plant and animal-originated waste management. It offers sustainable solutions to reduce and manage food waste, addressing issues like hunger, food loss, resource depletion, environmental challenges, and ensuring food security.
• AI and Food Quality: AI enhances food quality evaluation by assessing smell, appearance, texture, and taste. E-nose, E-tongue, and machine vision provide rapid and accurate diagnoses, improving quality indices, especially via aroma detection.
• AI and Food Safety: AI contributes to food supply and security by maintaining food quality and nutrients. It helps combat food fraud (driven by economic profit) through quick and correct diagnosis. AI models are crucial for predicting food risks related to climate change. It also helps manage the risks of hazardous chemicals and pesticides in agriculture, using ANN for pesticide reduction and E-tongue for evaluating hazard complexes.
• AI and Food Supply Chain: AI methods improve the supply chain, particularly for agricultural products. The massive increase in data size (due to population growth and climate change) makes AI efficient for supply chain management.
• AI and Food Crystallization: AI (mathematical/statistical algorithms) is a modern tool for predicting crystal behavior, perceiving and controlling crystallization, and simulating/discovering new crystals, enhancing a vital process in chemical industries.
• Smartphones, Software, and Practical Applications:
  • Smartphones with AI apps track food consumption, recommend diets, evaluate food quality/freshness via image analysis (benefiting manufacturers and consumers), and detect foodborne pathogens on-site.
  • IBM Food Trust: A blockchain-based platform using AI/ML for food traceability and safety from farm to table.
  • Blue River Technology: Uses computer vision and ML to spray weeds, reducing herbicides.
  • Agshift: Uses computer vision and ML on mobile phone images for automated food quality inspection, reducing waste from human error.
  • ImpactVision: Applies ML and hyperspectral imaging to predict food quality (e.g., freshness, ripeness).
  • ChatGPT (LLM): Can be used for crop forecasting, soil analysis, disease/pest identification, monitoring conditions (pH, nutrients, moisture), and generating alerts. Challenges include low data input quality and high cost.
• AI and Agriculture: AI is widely used to improve safety and quality of crops, soil, and weather conditions. It helps in seed choice, planting time, managing harmful factors (diseases, weeds, pests). ML and DL are popular for prediction and results using heterogeneous datasets (e.g., livestock health, dairy farm monitoring). ML is also popular for diagnosing diseases in agricultural products. Advantages include highest quantity/quality with minimum resources, reduced waste, and environmental harm.
• AI and Nutraceuticals, Human Nutrition: AI is effective in personalized diets, tracking food intake, understanding diet-disease relationships, food valuation via image processing, and lifestyle intervention. MLP (multi-layer perceptron) shows high accuracy for food classification and micronutrient analysis (calcium, protein, sodium, carbs, lipids, calories). AI applications act as personal assistants for diet modification. AI-based nutrition value analysis is gaining public interest due to high accuracy and rapidity.

6.2. Data Presentation (Tables)

The paper provides several tables summarizing characteristics of ML algorithms and specific applications.

The following are the results from Table 1 of the original paper:

Table 1 compares ANN, SVM, and DT based on performance items (Type of Task, Training Speed, Classification Speed, Dealing with Discrete/Binary/Continuous Attributes, Tolerance to Noise, Overall Accuracy) and datasets features (Size, Labeling/Non-Labeling). SVM generally shows High Overall Accuracy but is Not suitable for very large datasets, while ANN Requires large datasets with Moderate accuracy. DT Works well on large datasets but has Low accuracy, highlighting the trade-offs in algorithm selection.

The paper provides Figure 3, a schematic diagram illustrating various aspects of food assessment.

该图像是一个示意图，展示了食物评估的多方面内容，包括农业、分类、配方、感官评价、工艺、质量、安全、供应链管理、营养相关和废弃物处理等关键领域及其具体指标。

Figure 3: This diagram illustrates the interconnections and influence of various factors and processes within the food system, such as agriculture, categorization, formulation, sensory evaluation, processing, quality, safety, supply chain management, nutrition, and waste handling, indicating areas where AI can be applied.

The following are the results from Table 3 of the original paper:

Table 3 details the applications of E-nose, E-tongue, and machine vision across various fields (Categorizing, Quality, Food Safety, Sensory Evaluation) in FST. SVM is a prominent method for E-nose in classification and food safety tasks. PCA and LDA are frequently used alongside SVM for both E-nose and E-tongue in sensory evaluation and categorization. Machine vision primarily leverages MLP, CNN, and K-Mean for quality assessment and categorization of products like fish, nuts, alcoholic drinks, and saffron. This table vividly illustrates the versatility and impact of these AI-based devices in practical food applications.

The following are the results from Table 4 of the original paper:

Table 4 presents a detailed overview of AI and ML applications across various food product categories, including Dairy, Fermenting Related, Edible Oil, Cereal and Beans, Nuts, Beverage, and Nutraceutical and Functional Food. For each category, specific tasks (e.g., predicting demand, identifying products, detecting adulteration, forecasting yield, quality assessment) are listed alongside the algorithms employed (NN, DL, Machine vision, K-mean, FFANN, ANN, KNN, RF, SVM, LVQ, Extra Trees, PLS-DA, FT-Raman spectroscopy, ANFIS, 1D-CNN, 2D-CNN, PCA, XGBoost, MLP, BPNN, Mask-RCNN, CNN, BDT1, BDT2, RUSBT, DT, SWR, HCA, HSI system, FNN, CSA, ELM, NB, GBT, RF-GBT fusion, RSM). The conclusions highlight the effectiveness, accuracy, and potential of these AI methods in solving complex problems, improving efficiency, quality, and safety within these diverse food sectors. For example, RF and SVM show accurate classifying for Brazilian cheese, CNN provides authentic and reliable conclusion for lentil flour fraud, and DL offers great performance for pork meat water retention.

6.3. Ablation Studies / Parameter Analysis

The paper, as a review, does not present its own ablation studies or parameter analysis. These types of analyses are typically conducted in original research papers to evaluate the contribution of individual components of a proposed model or the impact of hyper-parameter tuning on performance.

However, the discussion within the paper often implicitly refers to such analyses conducted by the referenced works. For instance, when comparing linear discriminant analysis (LDA) and principal component analysis (PCA) for non-alcoholic beer classification, the paper states that LDA had a more accurate performance than PCA. This implies an underlying comparison or ablation-like study in the original research that differentiated the performance of these two methods. Similarly, the Conclusion section of Table 4 often highlights the accuracy or better performance of one algorithm over others (e.g., RF performance outweigh the other methods for coconut maturity grade, FFNN is better than SVM, LDA, LR, iPLS for sugar percentage forecasting), which are results typically derived from comparative analyses.

The paper also notes that the performance of ML is directly dependent on the type of its data, indicating an awareness of data characteristics as a parameter influencing results, though it doesn't present a specific parameter analysis conducted by the authors.

In summary, while the paper itself doesn't conduct these studies, it synthesizes the outcomes of such analyses from the literature to illustrate the effectiveness and nuanced performance of various AI/ML techniques in FSTI.

7. Conclusion & Reflections

7.1. Conclusion Summary

The paper unequivocally concludes that artificial intelligence (AI) is a critical advancement for the entire food cycle, from planting and crop cultivation to processing and consumption, and its impact on human health. By integrating AI, Internet of Things (IoT), and appropriate sensors and methods, it is feasible to establish an intelligent from-farm-to-fork cycle. This comprehensive system can manage the entire product lifecycle, encompassing agricultural steps, laboratory formulation, factory production and packaging, market marketing and retail, and ultimately, the physical and psychological effects on humans.

This intelligent system enables the tracking of mistakes or defects at any stage, making problem-solving more manageable and precise. The anticipated macroeconomic savings due to high accuracy and precise spot detection are substantial, largely by preventing many trials and errors. The paper emphasizes that AI's application extends to predicting climate change, crop yields, weather patterns, soil quality, optimizing planting methods, managing dairy farms, waste management and recycling, manufacturing processes for food products, functional foods, and nutraceuticals, sensory evaluation, nutrient content assessment, packaging, storage, supply chain management, food quality and safety assurance, distribution, consumer satisfaction, and personalized nutrition and diet management. It also highlights AI's potential in predicting and controlling food-related crystallization processes. In essence, AI and machine learning lay the foundation for a highly promising and intelligent cycle addressing the most fundamental human need: food.

7.2. Limitations & Future Work

The paper acknowledges several limitations and implicitly suggests areas for future work:

• Job Displacement: The most pressing societal concern identified is unemployment, as machines can replace many jobs, impacting employment standards. This is a significant socio-economic challenge that needs careful consideration in the widespread adoption of AI.

• Technological Challenges and Costs: AI implementation faces technological challenges, including the need for precise programming and the high costs of creating and maintaining AI systems. Constant updates and maintenance contribute to expenses, which could increase product prices and affect the affordability and accessibility of AI applications.

• Data Quality and Privacy: While non-traditional data sources (NDs) offer benefits, they also introduce challenges like biased data and privacy issues, especially in crowd-sourced surveillance for food safety. Researchers are actively working on these, but many solutions are still in the research phase. The paper mentions low data input quality as a challenge for ChatGPT leading to low prediction and analysis.

• Nascent Stage of AI Integration: The paper notes that the major utilization of AI in many food science and industry sub-domains is innovative and not yet routine. This implies a lack of research sources in many subgroups compared to other scientific subjects.

• Need for Further Research: The "lack of research sources" and the "novelty of a topic" present vast potential for numerous research avenues. The paper explicitly states that the subgroups of these two areas are countless, and the details of each project, even on the same subject, lead to a shift in the approach and ultimate goals. This suggests continuous, detailed research is needed to realize the full potential of AI in various niche areas of food science.

• Addressing Data Gaps: The need for large datasets for deep learning methods, and the quality of input data, are ongoing challenges that require future work in data collection, annotation, and management.

  Future work will likely focus on addressing these limitations, refining AI models for specific food-related tasks, developing more robust and privacy-preserving data handling mechanisms, and exploring the ethical and economic implications of widespread AI adoption in the food sector.

7.3. Personal Insights & Critique

This paper provides an excellent, broad overview of the current and potential applications of AI in the food domain, effectively positioning AI as a transformative force for the from-farm-to-fork cycle.

Inspirations and Transferability:

• Holistic Optimization: The concept of an intelligent from-farm-to-fork system is highly inspiring. It suggests that AI is not just a tool for isolated problems but an integrating technology capable of creating synergistic efficiencies across an entire value chain. This holistic approach could be transferred to other complex industries with long supply chains, such as pharmaceuticals or fashion, to optimize everything from raw material sourcing to consumer satisfaction.
• Multimodal Data Fusion: The emphasis on diverse data sources—from sensor data (E-nose, E-tongue, machine vision) to chemometric tools, agricultural parameters, and consumer analytics—highlights the power of AI in integrating and making sense of multimodal data. This is a critical lesson for any field struggling with disparate data streams.
• Personalization: The application of AI in personalized diets and nutraceuticals is a glimpse into a future where health and nutrition can be highly customized. This principle could extend to personalized medicine, education, or even tailored product development in other consumer goods sectors.

Potential Issues, Unverified Assumptions, or Areas for Improvement:

• Implementation Gap: While the paper extensively lists applications and benefits, the practical challenges of integrating these diverse AI solutions across a fragmented and often traditional food industry are immense. Many small to medium-sized enterprises (SMEs) in the food sector may lack the technical expertise, financial resources, or infrastructure to adopt these advanced AI solutions. The paper touches on high costs as a challenge, but a deeper analysis of socio-economic barriers and potential policy interventions would be valuable.

• Data Standardization and Interoperability: For a true from-farm-to-fork intelligent system, data standardization and interoperability across different platforms, sensors, and stakeholders are crucial. The paper mentions IoT for data transfer, but the underlying complexity of standardizing data formats and communication protocols across a vast ecosystem needs more emphasis.

• Ethical Considerations Beyond Employment: While job displacement is noted, broader ethical implications of AI in food, such as data bias leading to unfair agricultural practices or biased nutritional advice, food sovereignty concerns if AI is controlled by a few large entities, or the environmental footprint of AI computations, could be explored further. The paper mentions data bias as a challenge, but its ethical ramifications could be expanded.

• Validation and Reproducibility: As a review, the paper relies on the findings of others. A critical assessment of the validation methodologies and reproducibility of results in the cited AI applications would strengthen the overview, especially given the "low data input quality" challenge mentioned for ChatGPT.

• Human-in-the-Loop: The vision of fully automated AI systems is compelling, but the role of human experts (farmers, food scientists, nutritionists) in supervising, interpreting, and refining AI decisions remains critical. A discussion on optimal human-AI collaboration models would be beneficial.

  Overall, this paper serves as an excellent foundational text for anyone seeking to understand the current landscape and future trajectory of AI in the food sector. Its strength lies in its breadth and vision, offering a robust framework for future specialized research and practical implementation.