Skip to main content
Back to timeline
arXivSource publication:

Liquid food no longer needs a label: 2.4 GHz OFDM signals estimate carbohydrates, fat and protein directly

Synopsis

The carbohydrate, fat and protein content of liquid foods can now be estimated directly from 2.4 GHz OFDM radio signals passed through the liquid. Such quantification previously relied on package labels, laboratory analysis or manual user logging, while RF methods mostly identified liquid type or gave coarse nutrient fractions. ICOT first infers the ingredient classes that drive dielectric behavior, then regresses the three macronutrients; across 55 beverage classes and 28,420 captures, in-distribution MAE is 0.101 g/100 ml and leave-one-class-out MAE is 0.485 g/100 ml, though generalization to novel formulations remains the main open problem.

Source-provided article image: RF-Based Macronutrient Estimation in Liquid Foods via Composition-Aware Multi-Task Learning
arXiv

Interpretation

The 2.4 GHz OFDM channel response through a liquid can support quantitative estimation of carbohydrates, fat and protein, not just identification of which liquid it is. Earlier RF food sensing relied on microwave-oven leakage and reported coarse nutrient fractions rather than absolute per-volume quantities; vision and manual methods are limited by viewpoint, lighting, occlusion and the need for users to log intake. Across 55 beverage classes and 28,420 captures, in-distribution MAE is 0.101 g/100 ml and leave-one-class-out MAE is 0.485 g/100 ml, the latter comparable in magnitude to deviations in manufacturer-reported labels.

The model splits nutrient regression into two steps: first judging which dielectric ingredient classes dominate the liquid, then regressing nutrients conditioned on those classes. Direct regression from RF features is ambiguous because ionic conduction and non-ionic dipolar relaxation both appear as increased attenuation, yet correspond to non-caloric electrolytes and sugars respectively. The input is a two-dimensional channel frequency response matrix of 11 channels by 64 subcarriers, encoded by a Transformer encoder; ingredient labels supervise training only, and inference uses RF measurements alone.

Ingredient conditioning is functionally necessary in-distribution rather than an optional add-on. Shuffling the predicted ingredient embeddings at inference while keeping RF inputs fixed raises in-distribution error from 0.101 to 0.769 g/100 ml. The same shuffle barely changes leave-one-class-out error (0.485 versus 0.482 g/100 ml), indicating the conditioning does not transfer to unseen formulations.

Perspective

The results come from a controlled testbed: a fixed glass enclosure, a 30 mm path length, 2 L samples, a synchronized WARP v3 software-defined radio and directional antennas, making this a feasibility study rather than a deployable commodity sensor. The label-deviation comparison for in-distribution error applies only to in-distribution liquids.

Leave-one-class-out error is roughly four times the in-distribution error and concentrates in a few classes: 1% fat milk, the only unsweetened member of its fat regime, is misread as non-dairy, and high-sucrose dairy and tea mixtures have carbohydrates systematically underestimated. Temperature variation, mixture complexity and polar additives each add distribution shift, and generalization to high-concentration carbohydrates remains the key open problem.