Two different layers of data

Sensory evaluation and instrument measurement don't see the same layer of information. Sensory evaluation is a synthesis of smell, taste, mouthfeel, and attributes humans can perceive under an evaluation condition. Instruments can measure volatile compounds, absorbance, electronic signals, acidity, or many other physico-chemical quantities. So the goal of comparing them shouldn't be finding a machine that “tastes” exactly like a human, but understanding what each layer of data can and can't tell you.

Evidence of a relationship

Research by Lindinger and colleagues is an early example of using instrument data to predict the sensory profile of espresso. Studies comparing brewing methods also show analytical and sensory data can be placed side by side to describe differences between beverages. However, correlation only has meaning within a specific design. If you change the sample, brewing method, or evaluator without reconfirming, an old relationship isn't guaranteed to still apply.

Standardizing before comparing

The first step when combining two data sources is standardization. A sensory scale can have a completely different structure from an instrument signal. One variable might be measured in chemical units, while a sensory attribute is an intensity or acceptance score. You need to clearly define units, value ranges, how outliers are handled, and measurement conditions before looking for a correlation. Otherwise, a mathematical relationship can reflect a scaling artifact rather than an actual connection.

Validating the model

Validating a model matters just as much as building it. If all the data is used both to find a relationship and to prove that relationship, the risk of overestimating performance is very high. You should separate model-building data from validation data, or use a suitable confirmation design. This matters especially with sensory data, which usually has a limited sample size, making a complex model very prone to learning noise instead of signal.

Complementary roles of instruments and sensory evaluation

Instruments have a clear advantage in quantification and tracking certain chemical or physical variables with good repeatability. Sensory evaluation retains the integrated information of human experience that a single sensor can't fully represent. A 2025 review of objective methods in coffee evaluation also concludes in favor of combining the two: instrument-based and sensory methods each have strengths and limits.

So the practical question isn't “when can a machine replace a human taster,” but “which part of the process can a machine do better.” Instruments can screen, track trends, detect differences, or support explanation; sensory evaluation is still necessary for assessing the integrated experience. A good quality-control system should use instrument data to increase measurement capability and traceability, while still keeping sensory evaluation for what it does best.

When building a correlation model, you should start from a few sensory attributes with clear meaning rather than trying to predict an entire profile in a single model run. For example, one compound group might relate to one odor-attribute group within a given sample set. The model then needs to be tested on new samples. If the result only holds on training data, it hasn't yet proven the ability to replace or support evaluation.

Instrument data also needs to be collected under the same brewing conditions if comparing beverages. Research comparing nine extraction methods shows brewing method changes both analytical data and sensory profile. So if you mix data from different brewing methods without including method as a model variable, the relationship found can be confounded. Standardizing brewing conditions is part of data standardization, not a separate step.

Another risk is confusing discriminative ability with predictive ability. An instrument can distinguish two samples very well but still fail to predict a sensory score for a new sample. To talk about prediction, you need independent validation data. If the goal is only screening, evaluation criteria can differ. Defining the goal from the start helps choose the right way to assess the model.

Sensory evaluation also has its own variation, so sensory data needs to be standardized for water, ratio, temperature, time, and sample-presentation method according to the research procedure. If the sensory layer isn't stable, the instrument model will be evaluated against a noisy target. So model quality depends on the quality of the sensory reference data too, not just the quality of the instrument.

A good dataset should keep a one-to-one link between sample code, brewing condition, instrument data, and sensory scorecard. When you need to recheck a correlation, you can trace back to the exact sample and original conditions. If only summary tables are kept and the linking codes are lost, detecting errors or validating the model becomes much harder. Traceability is therefore part of data quality, not just a management requirement.

References

• Lindinger et al. (2008), When machine tastes coffee: instrument-based approach to predict the sensory profile of espresso coffee, Analytical Chemistry 80, 1574–1581, DOI 10.1021/ac702196z.
• Gloess et al. (2013), Comparison of nine common coffee extraction methods: instrumental and sensory analysis, European Food Research and Technology 236, 607–627, DOI 10.1007/s00217-013-1917-x.
• The prospects of replacing the sensory coffee quality assessment methods with objective, instrument-based approaches: A review (2025), Trends in Food Science & Technology 163, 105210, DOI 10.1016/j.tifs.2025.105210.
• Hong et al. (2024), Food Chemistry: X 21, 101119.