2025 AIChE Annual Meeting

Predictive Modeling to Measure Chemical Decay in Roasted Coffee

Freshness is a critical quality parameter in coffee, yet it remains defined by subjective timelines such as “best by” dates rather than by measurable chemical and physical properties. From a chemical engineering standpoint, roasted coffee behaves as a reactive porous solid undergoing mass transfer and reaction processes. CO2 diffusion, volatile organic compound degradation, and lipid oxidation collectively drive staling, altering both flavor and extraction efficiency. Current industrial practices fail to capture these dynamics, often leading to premature product disposal and waste of a resource-intensive agricultural product.

This research develops a quantitative framework for assessing coffee freshness by integrating multi-experimental data with machine learning. Roasted beans were aged under controlled storage conditions at multiple roast levels, and changes were monitored across gas-phase (CO2 release, volatile degradation using GC-MS), solid-phase (lipid oxidation using FTIR-DRIFTS, mechanical hardness), and sensory analyses.

These datasets were analyzed using machine learning models, including random forest regression, support vector regression, gradient boosting, and deep learning approaches such as feed-forward neural networks. Hyperparameter tuning was conducted using grid search and cross-validation to optimize predictive accuracy. To address limited experimental data and improve generalizability, a generative adversarial network (GAN) was trained to produce synthetic feature distributions that expanded the dataset while preserving underlying statistical relationships. Incorporating this augmented dataset improved model robustness and reduced overfitting, enabling accurate prediction (R² > 0.9) of coffee age across roast levels and storage times. Feature importance analysis highlighted CO2 and 2-Methylfuran content decay as the most important predictive features. The integration of experimental data and machine learning enabled systematic identification of compound-specific degradation pathways most relevant to perceived freshness.

This study demonstrates how chemical computational methods can change our understanding of freshness from a subjective metric into a variable that can be modeled and controlled. By quantifying freshness in predictive models, this work provides a foundation for evidence-based process control strategies in coffee roasting. More broadly, the methodology illustrates how data-driven approaches can be applied to food and bioproduct systems, combining chemical and sensory data under a predictive model.