How Science-Based AI Enables Supply Chain Resilience

When new formulations are required, science-based AI can help development teams quickly zero in on alternative ingredients, saving time in the lab so the business can keep the products that consumers want on the shelves.

Generative Art Adobe Stock 609457865
Generative ART AdobeStock_609457865

Food companies are no stranger to supply chain challenges, especially in recent years as the COVID-19 pandemic upended supply routes and the availability of ingredients worldwide. The disruption of the pandemic imparted some valuable lessons, however, as more and more companies have redoubled their focus on being nimble enough to change course quickly when supply chain issues arise.

But fast pivots are easier said than done, especially when it comes to the complex formulations that many food companies rely on to ensure quality, flavor and performance. When a critical formulation ingredient becomes suddenly unavailable, finding an alternative is no simple task. A great deal of time, money and R&D resources often go into developing formulations. It’s research that can take months or even years, requiring massive amounts of data, complex chemistry and time-consuming lab-based experiments. 

Advancements in artificial intelligence (AI) technology focused on the challenges of chemical and materials R&D are opening the path to a better way. Imagine having the ability to explore an exponentially greater field of formulation ingredients in-silico, screening for specific product and supply chain requirements, and identifying viable candidates for experimentation, all in a matter of minutes vs. months. This is the promise of science-based AI.

Let’s take a closer look at the challenges faced by today’s food companies when it comes to formulation development and at how science-based AI can help: 

Traditional formulation development is too slow 

When faced with an ingredient shortage, price change, recall or new regulation, product development teams are under an enormous amount of pressure to come up with alternative formulations within very narrow timeframes. But this time pressure is oftentimes at odds with the scientific complexity involved in identifying, formulating and testing alternatives. Scientists working to develop food formulations must account for a multitude of parameters, including chemical variables that impact flavor, smell, texture, shelf stability and more, as well as health, nutritional and regulatory concerns. Traditionally, this work has been done through a combination of lab experimentation, which is time consuming and costly, as well as traditional simulations, which is also slow and requires a lot of human analysis, at multiple scales, to yield insights and predictions. 

AI offers promise, but classical machine learning is not enough

The type of AI that everyone is currently talking about in the news (i.e. ChatGPT) is based on classical machine learning. This method is all about compiling a lot of data to extract insights. The difficulty in using classical machine learning in product development is that most problems facing formulation scientists don’t have the advantage of massive datasets. And if researchers use big generic datasets (rather than their own proprietary data, or scientifically relevant data) to train their models, the results are often not accurate enough to be useful, lacking any grounding in scientific laws or knowledge.       

Science-based AI = Machine learning + scientific rigor  

Science-based AI merges the scientific rigor of lab experimentation and traditional simulation with machine learning. Much like how scientific theory allows for extrapolation from experimental data, science-based AI integrates physical laws, chemical properties and system constraints into machine learning models that have the ability to address complex formulation problems using sparse data sets. This approach is especially appealing for unexpected challenges like a supply chain shortage, because scientists can use “inverse design” to create models that, instead of using a set of inputs to predict an output, start with an output. For example, one of the ingredients in an emulsifier or preservative is suddenly unavailable. Researchers can use an inverse design model to quickly find alternate ingredients that work in the formulation without impacting shelf-life or taste.    

When an AI model is trained with scientific principles, understanding of chemical systems and relevant data sets, accuracy is improved and product development teams then can run thousands of experiments virtually, saving expensive and time-consuming lab experiments for the most promising candidates. Science-based AI models can also help teams evaluate the trade-offs between different alternatives in terms of cost, availability, safety, nutritional value or environmental impact. And all of this can be done much more quickly than with traditional lab and simulation-based development, shrinking reformulation cycle times significantly.   

Food companies aspiring to use AI to pave the way for faster formulation should break the process down by first determining their requirements, identifying relevant data, and using a science-based AI model that understands the system requirements, the data, and the relevant scientific principles.  

  • Determining the requirements for a specific formulation challenge may encompass exploring how alternative ingredients impact the chemistry of the final product’s stability, shelf-life, flavor, scent, texture or nutritional profile.   
  • Relevant data can come from both internal and external sources. There are expansive open source databases that cover a broad design space of chemical compounds. But this type of data must then be amplified with internal data that’s specifically relevant to your product. 
  • Combining the system requirements, data from both internal and external sources, along with scientific laws and knowledge marks the boundary where traditional, data-driven AI ends, and science-based AI commences. And the beauty of science-based AI is that science-infused machine learning models can reach much higher levels of accuracy and deliver deeper insights with far less data and training time than classical machine learning systems. This empowers scientists with information they need to confidently make critical product formulation decisions, faster.

In a volatile world, food companies need to be able to adapt quickly to supply chain challenges that impact the availability of critical product ingredients. When new formulations are required, science-based AI can help development teams quickly zero in on alternative ingredients, saving time in the lab so the business can keep the products that consumers want on the shelves.