Traditional approaches towards creating new foods are too slow
Creating innovative food products traditionally involves a complex process that combines food science, engineering, culinary art, and consumer research and heavily relies on iterative cycles of gradual improvement. Figure 1 illustrates the food development cycle for the example of a new plant-based meat product: The first step in the design process is to precisely define the desired product by identifying the target meat, e.g., chicken, pork, or beef; selecting the specific cut, e.g., burger, sausage, or steak; establishing key features, e.g., texture, flavor, appearance, or nutritional profile; and understanding consumer preferences, e.g., allergies, dietary restrictions, or environmental concerns. The second step is to select the ingredients by choosing protein sources, e.g., soy, wheat gluten, pea, or bean, to deliver the desired structure and nutritional profile; choosing fats and oils, e.g., coconut oil, canola oil, or shea butter, to mimic juiciness and mouthfeel; incorporating binders, e.g., methylcellulose or starch, and functional additives, e.g., carrageenan or lecitin, to enhance texture, binding, and stability; and adding flavors, e.g., yeast extracts or fermentation-derived compounds, to replicate umami or meaty flavors. The third step is to develop the formulation by optimizing the ratios of proteins, fats, binders, and additives to achieve the desired sensory attributes; addressing nutritional needs, e.g., high in protein, low in saturated fat, fortified with vitamins or minerals; and integrating flavor compounds to fine tune the taste, e.g., meaty, fatty, or smoky. The fourth step is to engineer the texture by choosing the processing method, e.g., extrusion, spinning, or 3D printing, to replicate the fibrous, layered structure of animal muscle; optimizing rheological properties, e.g., tensile, compression, or shear strength, to mimic the resistance to chewing; and designing a fat and moisture retention system to create the juiciness of the target product. The final step is to optimize the product, for example, by adding colorants, e.g., beet juice, annatto, or paprika, to improve product appearance; by performing customer surveys to satisfy texture and flavor preferences; and by improving product safety and shelf stability. By its very nature, this traditional approach involves dozens of cycles to develop formulations, probe texture, prepare samples, and survey consumers. During these iterations, a change to any of the parameters in any of these steps can result in significant variations in the final product, which are often highly unpredictable. Obviously, this trial-and-error approach is time-consuming, expensive, and inefficient, especially when considering the urgency to transform our current food system. But fortunately, any of these steps provides an opportunity for AI: AI can drive ingredient selection, formulation development, texture engineering, and product optimization, and efficiently screen a massive multimodal parameter space to identify the most promising parameter combinations.
Before we answer this question, it is important to understand that there are two different types of AI, non-generative and generative AI: non-generative AI analyses, improves, or infers data without creating new data, whereas generative AI creates new data that resembles existing data. Three traditional applications fall into the category of non-generative AI: optimization, probably the most widely used application of AI for food today, where the AI fine-tunes variables to achieve best possible outcome under certain constraints, for example, by optimizing ingredient combinations to maximize nutritional value and minimize environmental impact; discovery, where the AI finds insights, patterns, and trends from data, for example, by identifying new protein sources from analyzing the chemical and mechanical properties of various plants to determine their suitability for mimicking the texture and taste of animal meat; and prediction, where the AI forecasts outcomes or behaviors, for example, by predicting the taste of a combination of ingredients or the preference of consumers towards novel alternative protein products. One very recent application falls into the category of generative AI: creation, where the AI generates entirely new ideas, formulations, or textures, for example, by creating entirely new formulations only on the basis of natural language prompts.
To answer this question, let's look at the example of replacing animal meat by an alternative protein product: The objective is to discover a new formulation for a product that either satisfies desired properties, e.g., nutrition, texture, or flavor, or mimics a specific target product, e.g., an existing, resource-intensive animal product that we seek to replace. The new product also needs to satisfy certain constraints such as nutritional profiles, texture, and flavor. We may also want to include or exclude water or specific ingredients, for example, to modulate texture or address food allergies. And we may want to include additional regional, seasonal, or environmental constraints. From this input, the AI would create a set of new formulations as output, where each formulation consists of a list of ingredients with their respective fractions or weights. The AI could further optimize these formulations, for example, by constraining the number of ingredients, or reducing their environmental impact or cost. In addition, the AI could also optimize an associated set of process parameters, for example, extrusion velocity and pressure, cutting, cooling, or heating, to achieve a desired texture and rheology. From the optimized formulation-encoded through the weighted ingredient list-the AI could predict properties, for example, the nutritional profile of the final product.
While nutritional profiles are relatively easy to predict from a list of weighted ingredients, it is a lot more challenging to predict rheology, texture, or flavor. This is not a general limitation of AI as a technology per se; rather, it is a temporary limitation that reflects the current lack of appropriate data or our inability to process big data at scale. Using AI to generate new foods is still in its infancy, and data that correlate formulation to rheology, texture, and flavor are rare. Labeled and structured data are often proprietary, as they require significant time, expertise, and resources to generate-especially in food science, where expert annotation adds substantial value. Yet, only few approaches in the literature use unsupervised learning or reinforcement learning, while most AI technologies for food today still heavily rely on supervised learning based on labeled data and human feedback: Food scientists pilot production using the new formulation and process parameters, engineers probe its rheology and texture, chefs prepare it for sensory surveys, and consumers taste and annotate it for taste, flavor, texture, and overall customer satisfaction, similar to the graphic in Fig. 1. This laborious process might not immediately result in the most optimal product. Nonetheless, all the steps do naturally generate new training data that will provide useful information when creating future products. Ingredients, formulations, nutritional profiles, rheology, texture, flavor, and taste could constitute valuable data for a foundation model, a large pre-trained multimodal model that understands relationships between these variables. Similar to many other applications of AI, we could envision a division of labor, where the process of building and pre-training the foundation model is performed by data science specialists while food science specialists would fine-tune the model to their specific needs.
