From Data to Finished Product

Abstract

Richard Marshall gives an overview on AI and how it is transforming food product development in the UK, enhancing everything from recipe optimisation data-driven innovation. Advanced techniques such as inverse design allow food scientists to create new products by working backwards from desired characteristics, accelerating development and boosting success rates.

Artificial Intelligence (AI) is playing an increasingly significant role in both personal lives and in all areas of industry. Many people will have already used it perhaps unknowingly, for example when using an internet search engine or using automotive parking assistance. AI includes a number of different levels of data manipulation. It forms the basis of expert systems that analyse complex data to produce results. At the next level, machine learning (ML) algorithms learn from data to make predictions or decisions. A more advanced version of machine learning has artificial neural networks (ANN) which mimic human brain function, taking data in, analysing it in some form of ‘black box’ and then presenting the results. More recently, we have seen the development of large language models (LLM) that use natural language inputs and outputs – they ‘understand’ discursive questions. Within such systems, data is often analysed using fuzzy logic when it has varying levels of ‘truth’. AI is very suitable for use in product development particularly using inverse design, that is starting from knowing about products, recipes and ingredients.

Artificial intelligence (AI) is the simulation of human intelligence in machines. It involves the development of methods and software that enable computers to perceive their environment and react to it. AI aims to mimic human thinking and problem-solving abilities⁽¹⁾.

AI is becoming more and more prominent in our lives. It has the potential to give significant benefits to the way we live, to make industry more competitive and more efficient in ways that we never imagined a few years ago. As with any emerging technology, there are risks and fears as well as positive opportunities. The UK Government has recognised this and hosted an international conference on AI in 2023⁽²⁾. We are generally unaware that AI is already playing a major role in many activities. Internet search engines, such as Google, use it to present results to users, often in a way that biases towards favouring advertisers, promoting certain views or suppressing certain sites⁽³⁾. The Internet of Things (IoT) enables smart devices, such as domestic fridges, smart watches and autonomous vehicles to communicate with the world, sharing data, providing assistance and information. The food industry is no exception in this regard. Robotics has been used for some considerable time to move product around factories, control process operations through automation and optimising supply chain management⁽⁴⁾.

At a basic level, expert systems use a form of AI in which users feed information into a human-machine interface, i.e. a computer keyboard, touch screen etc⁽⁴⁾. The initial knowledge for expert systems is acquired from a human expert. Within the system, an inference engine compares the input from the interface with rules in the knowledge base. The knowledge is in the form IF (condition)….THEN (conclusion) followed by logic operations AND, OR, NOT. Inside the system, data is held during the operations until a result has been generated which is then presented to the operator via an interpreter module, which may display on a screen, give a signal or print out. Expert systems are already widely used across the food industry (Table 1).

ML and ANNs can use ‘fuzzy logic’ where the ‘truth’ value of variables can range from 0 (not true) to 1 (completely true). The ‘truth’ can be thought of as a weighting that describes how true a value is. In standard Boolean logic, variables are either true or false, 0 or 1. Fuzzy logic allows for uncertainty in variable values, enabling the processing of imprecise data.

In a fuzzy logic system, sharp data (i.e. data that has real numbers) is transformed by a ‘fuzzifier’ into fuzzy data that indicates how reliable the values are. The fuzzy data input set is used to infer the meaning of the data based on a set of rules. Once this has been generated, the fuzzy output set is passed through a ‘de-fuzzifier’ that generates sharp or crisp output that can be used by the operator.

Examples of the use of fuzzy logic systems include modelling food control, to classify products and in handling fresh produce. The advantage of fuzzy systems is that they can use natural language for processing and they are good at managing multivariables and non-linear situations. For example, such a system can use sensory terms such as ‘not satisfactory’, ‘fair’, ‘medium’, ‘good’, and ‘excellent’ and extract useful information⁽⁶⁾.

Managing food safety is complex: it needs extensive knowledge and experience. Food safety management systems cover processing, monitoring, testing, training and maintenance. Safety relies on HACCP to evaluate hazards and the risk of occurrence. Safety is only improved after hazardous events through the evaluation of ‘lagging indicators’. The use of AI enables a more proactive approach using ‘leading indicators’. Such a system is not stand-alone but must work in conjunction with food safety experts. Examples of the challenge from microbiology facing the use of AI can be illustrated by considering that Salmonella has over 2500 serovars and Listeria monocytogenes has as many strain-dependent virulence factors⁽⁷⁾. According to these authors, AI is being used for rapid monitoring of microbial contamination in chicken liver meat, rapid verification of fresh produce wash water sanitation, predicting foodborne disease outbreaks, and predicting food safety compliance in food outlets.

As shown by Flynn (2023)⁽⁸⁾ in this journal, there are a number of applications for preventing food recalls or managing food safety systems or identifying fish species. On the farm, an AI-based system called Chirrup (chirrup.ai) can be used to identify bird species and so indicate the level of biodiversity.

LLMs are based on ANNs. They are able to communicate with users via natural languages and work by learning statistical relationships between words in text. They make use of fuzzy logic to generate output. Development started from about 2017 with Google being one of the first developers. They introduced Generative Pre-trained Transformer 1 (GPT-1) in 2018 followed by GPT-2 (2019), GPT-3 (2020) and GPT-4 (2023). Access is limited to GPT because of fears of abuse but ChatGPT is the free, publicly-available version, however this uses data from GPT-3.5, which only goes up to 2022. A pay-to-use version, ChatGPT Plus, is up-to-date.

As Flynn (2023)⁽⁸⁾ indicated, ChatGPT and similar LLMs can be interrogated to find information on food topics. This is equally true for food product development (FPD). However, the free-to-use ChatGPT cannot give up-to-date information on, say, market trends in FPD as its database is about two years old. On the other hand, asking ‘How can I use ChatGPT for food product development’ gives a help list that can act as a starting point (Table 4). This would be particularly useful for food science and technology students planning to develop a product as part of their coursework.

If a paid version of LLM is used, e.g. ChatGPT Plus, more current information can be obtained. It can provide trend spotting, discover new ingredients and technologies, emerging flavours, help choosing the best/right ingredients, choosing the process conditions, e.g. cooking times and temperatures, improving sustainability and carbon footprint and overall market trends and consumer preferences⁽⁹⁾.

AI can be very effective in supporting product innovation from idea generation, through building the business case, to the design, engineering, development and testing of the new product⁽⁹⁾. In the initial stages of new product development (NPD), LLMs can be used to generate novel ideas by scanning the internet, finding market opportunities from sources such as blogs, forums, reports, complaint lines and comments from product users. One area that LLMs excel in analysing is discursive data from surveys. They can generate the first draft of a customer interview guide. With information from these areas, LLMs can set out and evaluate concepts and link these to R & D data held by a company, exploiting intellectual property (IP).

With these first steps completed, LLMs can then be used to build the business case for the development of new product. This includes evaluating market data, predicting potential sales, learning about competitors’ activities and projecting income and profit. The AI can even prepare the business case for the NPD to be presented to company management.

After the go-ahead decision, AI has a further role in creating virtual prototypes, setting design parameters, defining development iterations and optimising the process. Initial product evaluation by test panels of consumers can be fed back into the design process with AI indicating further improvements.

Examples of recent applications of AI in food design include using ML models to predict sensory scores of chocolate cookie recipes⁽¹⁰⁾ and to identify optimal conditions for the ideal yoghurt sauce considering 36 combinations and 22 sensory attributes. ML models have also been used to predict co-occurrences of ingredients in recipes to discover novel food pairings. It has been noted on various occasions that new food products fail in the market place up to 75% of the time⁽¹¹⁾. Al-Sarayeh et al. (2023)⁽¹⁰⁾ argue that understanding consumer demands and using these to design a food product with appropriate chemical and physical characteristics is challenging for the food industry. They propose a different way of developing new products called ‘inverse design’. In conventional NPD, sensory, nutrition, health, convenience, psychology and social information are used to drive the creation of a new concept. One or more prototypes may be produced in an iterative process requiring multidisciplinary knowledge. After several development and processing trials, a product is produced. This complex process, involving many factors, is often costly in terms of both time and money, with potential for failure at any stage.

‘Inverse design’ begins by identifying the desired functionalities of a product and works towards optimising the design. In order to do this effectively, the relationships between inputs and outputs have to be modelled, which is too complex for humans to model manually. The application of AI enables more efficient and more effective development of the new product. Information from public recipe databases, food composition, structural and molecular properties can be fed via an encoder to the virtual design space. In that space, the AI can find novel products with targeted attributes. The results pass out through a decoder and are presented as a recipe for the new product with details of processing required, the sensory properties, nutrition and health aspects. These properties can then be measured to see if the product meets the specification. Consumer responses can also be used to assess the design. Al-Sarayeh et al. (2023)⁽¹⁰⁾ give an example from Marin et al. (2019)⁽¹²⁾ in which the ingredients of a recipe, the processing directions and visual properties are encoded by an AI deep learning model.