Algorithm-driven sweetener blending platform for sucrose-like sensory profiles
Machine learning platform that predicts natural sweetener blends with superior taste. Uses reinforcement algorithms to reduce off-notes like bitterness and lingering sweetness, producing formulations that closely match sucrose. Validated through iterative sensory panel testing across beverages, baked goods, and ice cream applications.
Overview
This research offers an algorithm-driven approach to developing natural sweetener blends that more closely replicate the sensory profile of sucrose. Natural sweeteners such as stevia often carry off-flavors including bitterness, lingering sweetness, and metallic notes that limit their commercial appeal. By applying machine learning and reinforcement algorithms, the platform predicts optimal combinations of natural sweetening agents, then validates them through human sensory panels in an iterative feedback loop. The result is a data-driven method for designing sweetener blends that reduce undesirable off-notes while delivering a taste experience consumers prefer.
Technical specifications
Core approach:
- Reinforcement learning algorithms adapted from fuel blending optimization applied to sweetener formulation
- Binary mixture sensory data feeds predictive models to identify optimal multi-component blends
- Iterative cycle of algorithmic prediction, sensory panel evaluation, and model refinement
- Parallel sensory methodology using overall-difference-from-control testing for holistic comparison to sucrose
- Validation across multiple food categories: cold beverages, baked applications, and ice cream formulations
Key benefits for partners:
- Reduces or eliminates off-notes associated with individual natural sweeteners
- Supports cleaner-label formulations using natural sweetening agents
- Accelerates product development by predicting high-performing blends before physical testing
- Applicable across diverse food and beverage categories with different temperature and processing requirements
Technology readiness level
The platform has been validated through initial rounds of sensory testing, with results demonstrating progressive improvement in blend performance as the algorithm learns. Data confirm that blending reduces off-notes in common natural sweetening agents and that the algorithms can predict better-tasting blends using prior sensory data. Future validation includes several additional rounds of sensory testing and predictive modeling, parallel difference-from-control testing, and consumer testing across cold beverages, baked goods, and ice cream to ensure robust performance across categories and temperatures.
About Cornell University
Cornell University is a comprehensive private, land-grant research university with campuses in Ithaca and New York City, combining significant scale with cross-disciplinary breadth. Industry connects through open-access user facilities and prototyping labs, pilot-scale testbeds, and a research and technology park that provide pathways from discovery to demonstration. A statewide extension network and integration with a major hospital system enable real-world deployment, while a graduate campus embedded in New York City’s tech corridor provides direct access to startups, venture investors, and corporate R&D teams. Research is supported by competitive federal funding from agencies such as the National Science Foundation, National Institutes of Health, the Department of Energy, and the U.S. Department of Agriculture. A dedicated technology transfer office streamlines IP management, licensing, startup formation, and corporate partnerships across campuses.
delivered to your inbox.