Have you ever bitten into a nut or piece of chocolate, expecting a smooth, rich flavor, only to be met with an unexpected chalky or sour taste? That taste represents food spoilage in action and affects almost every product in […]
Have you ever bitten into a nut or piece of chocolate, expecting a smooth, rich flavor, only to be met with an unexpected chalky or sour taste? That taste represents food spoilage in action and affects almost every product in your pantry. Now, artificial intelligence (AI) can help scientists tackle this problem with more precision and efficiency.
We are a group of chemists studying ways to extend the shelf life of food, including those prone to spoilage. Recently, we published a study describing the benefits of AI tools in preserving the freshness of oil and fat samples. Since oils and fats are common ingredients in many types of food, including chips, chocolate, and nuts, the study’s results can have broad applications and impact other industries, such as cosmetics and pharmaceuticals.
Food can become acidic when exposed to air over a certain period of time – a process known as oxidation. In fact, many common ingredients, but especially lipids (fats and oils), react with oxygen. The presence of heat or UV light can accelerate this process.
Oxidation leads to the formation of smaller molecules, such as ketones and aldehydes, which give a sour smell to spoiled food. Consuming spoiled food can jeopardize your health.
Fortunately, both nature and the food industry have excellent protection against food spoilage: antioxidants. Antioxidants encompass a wide range of natural molecules, such as vitamin C, and synthetic molecules that can protect your food from oxidation.
While there are several ways antioxidants work, they generally neutralize some of the processes that cause food spoilage and preserve the taste and nutritional value of food for longer periods. Many consumers may not even be aware that they are consuming added antioxidants because food manufacturers usually add them in small amounts during preparation.
However, you can’t simply sprinkle vitamin C on food and expect to see a preservative effect. Scientists must carefully select specific sets of antioxidants and calculate the precise amount of each.
Combining antioxidants does not always enhance their effects. In fact, there are cases where using the wrong antioxidants or mixing them in incorrect ratios can reduce their protective effects – this is called “antagonism.” Finding combinations that work for specific types of food requires numerous experiments that take time, require specialized staff, and increase overall food costs.
Exploring all possible combinations would require a tremendous amount of time and resources, so researchers are limited to a few mixtures that provide only a certain level of protection against spoilage. This is where artificial intelligence comes into play.
You have probably already heard of AI tools like Chatbots in the news or even played with them yourself. These types of systems can process large amounts of data, identify patterns, and generate output that can be useful to the user.
As chemists, we wanted to teach an AI tool how to search for new combinations of antioxidants. To do this, we chose a type of AI that can work with textual representations – written codes describing the chemical structure of antioxidants. We first provided our AI with a list of about one million chemical reactions and taught the program some basic chemical concepts, such as recognizing important features of molecules.
Once the machine was able to recognize general chemical patterns, such as ways certain molecules react with each other, we refined its training by teaching it more advanced chemistry. For this step, our team used a database of about 1,100 mixtures previously described in scientific literature.
At this point, the AI could predict the effect of combining any set of two or three antioxidants in less than a second. Its predictions matched the described effects in the literature in 90% of cases.
However, these predictions did not completely align with the experiments our team conducted in the laboratory. In fact, we discovered that our AI could accurately predict only a few oxidation experiments we performed with real pork fat, indicating the complexity of transferring results from the computer to the lab.
Fortunately, AI models are not static tools with predefined paths of yes or no. They are dynamic learners, allowing our research group to continue feeding them with new data until they perfect their predictive capabilities and accurately predict the effect of each combination of antioxidants. The more data the model receives, the more accurate it becomes, just like humans grow through learning.
We found that adding around 200 examples from our lab enabled the AI to learn enough chemistry to predict the outcomes of experiments conducted by our team, with only a small difference between the predicted and actual values.
Models like ours may one day help scientists develop better ways of preserving food by creating the optimal combinations of antioxidants for specific foods they work with – as if they have a highly intelligent assistant.