Food System Innovations has launched the Food Intelligence Lab, a new AI research initiative backed by a $2 million grant from the Bezos Earth Fund. The lab will build open-source datasets, machine learning models, and benchmarking tools designed to shorten product development timelines and improve the taste and texture of plant-based and other sustainable proteins.
The launch arrives as alternative protein companies still struggle with consumer adoption, often because products fall short on sensory performance despite years of formulation work. The lab's core bet is that shared data infrastructure can make AI-driven formulation as routine as it has become in drug discovery and materials science.
Anna Thomas, director of machine learning at the Food Intelligence Lab and a computer scientist at Stanford University, said: "AI is already transforming fields like drug and materials discovery, but food still lacks the shared infrastructure needed to fully unlock the potential of AI in this space. We're building tools to help food scientists iterate faster and create truly exceptional sustainable protein products."
Unlike pharmaceuticals, where large public datasets have sped up AI development, food formulation is slowed by fragmented data, proprietary research, and expensive experimental cycles. The lab aims to fix that by creating large-scale, open datasets that pair sensory evaluations with instrumental measurements like texture profile analysis, pH, and shear testing. These resources will feed AI models that predict consumer-relevant attributes before physical testing begins. The initiative will also work with food companies, academic researchers, and nonprofits to embed the models into commercial R&D workflows.
Closing the data gap in food formulation
For product developers, the bottleneck often isn't a lack of ideas but how long it takes to test them. A single formulation tweak can require weeks of internal panels or external consumer testing. The Food Intelligence Lab's open datasets are meant to give scientists a faster, data-rich starting point, reducing the number of physical prototypes needed and cutting time to final formulation.
The lab plans to release benchmarks and models that anyone can use, lowering the barrier for startups and smaller manufacturers that can't afford dedicated data science teams. By combining scientific literature, experimental records, and foundation models, the system should eventually recommend the next best experiment - a capability that could trim months from development cycles.
Early results show faster optimization
One early collaboration with Proxy Foods AI demonstrates the approach. The teams built an optimization system called Expert-Guided Bayesian Optimisation (EGBO) and applied it to a plant-based Greek-style yogurt. Over just five days and 10 formulation iterations, the AI system improved the yogurt's sensory performance by 29%, matching an animal-based benchmark on consistency, creaminess, and tanginess. It also outperformed a professional food scientist working under the same time constraints, reaching a higher optimization score faster.
Panos Kostopoulos, founder and CEO of Proxy Foods AI, said: "Food scientists shouldn't have to spend months on trial-and-error to get texture, mouthfeel, flavour, and aftertaste right."
He added: "Partnering with FSI's Food Intelligence Lab to open-source these tools is how we accelerate those breakthroughs and ultimately change how we feed the planet for the better."
Predicting what consumers will taste
Beyond formulation optimization, the lab is developing AI tools that predict sensory outcomes. One early output is TasteBench, an open benchmark and Kaggle competition that tests how well models can predict how closely a sustainable protein product resembles its animal-based counterpart. The strongest current AI model performs at roughly the level of the median human sensory panellist.
For R&D teams, that means a predictive tool that can screen formulations before committing to consumer panels, which are often the most expensive and logistically demanding part of product development. As the models improve, the number of physical panels could shrink, freeing budget for more creative exploration or faster scaling.
Why this matters for product development
Product developers in alternative proteins face a tight loop: speed matters, but sensory quality makes or breaks adoption. The Food Intelligence Lab's open tools offer several concrete shifts for daily R&D work:
- Instead of months of trial and error, teams can run hundreds of virtual formulation tests in days, then prototype only the most promising candidates.
- Open-source benchmarks and datasets mean smaller companies can access the same predictive capabilities as large manufacturers, reducing the cost of entry to high-quality formulation.
- AI that predicts taste and texture at the level of a trained human panellist could cut reliance on expensive sensory panels, speeding up decision-making and reducing per-product development costs.
- The EGBO system shows that guided AI optimization can outperform a skilled human formulator under time pressure - a direct signal that integrating these tools into standard R&D workflows can boost output without adding headcount.
For teams asked to deliver better-tasting products on tighter deadlines, the lab's data and models represent a practical path to fewer bench hours and more informed formulation decisions.
Your membership also unlocks:
