Coffee Roast Design of Experiments for Repeatability

ISEF Category: Engineering Technology: Statics and Dynamics

Subcategory: Industrial Engineering-Processing  ·  Difficulty: Advanced  ·  Setup: School Lab  ·  Time: Full Year

The Hook

Coffee roasting looks simple until two batches with the same settings taste different. That is a process control problem, not just a cooking problem. You can treat roast quality like a manufacturing line and measure how closely you hit a target profile. That makes this a strong science fair topic with real data, not just taste opinions.

What Is It?

This project asks a simple question with a hard answer. How do roast temperature, time, airflow, and batch mass change the final coffee profile? You are not just making coffee, you are studying a process. In manufacturing, that kind of study is called design of experiments, or DOE. DOE helps you test several factors at once instead of changing one thing at a time forever.

Your outputs can include total dissolved solids, or TDS, which estimates how much material moved from the grounds into the brew. You can also measure aroma with a sensor array, sometimes called an electronic nose, which reads patterns from gas sensors instead of using your nose alone. Then you can train a simple machine learning model to predict roast level from those measurements. Think of it like teaching a computer to spot the difference between light, medium, and dark roast from a signal pattern.

Why This Is a Good Topic

This topic works well because every part can be measured, compared, and improved. You have clear inputs, clear outputs, and a real goal, which is to hit a target roast with fewer trial runs. That connects to industrial process control, food engineering, and quality engineering. You can learn DOE, calibration, data cleaning, and basic ML in one project.

Research Questions

• How does drum temperature change TDS, aroma sensor response, and predicted roast level??
• How does airflow change the consistency of roast outcomes across repeated batches??
• What is the effect of batch mass on the spread of TDS values for the same roast setting??
• To what extent does a two-factor DOE explain roast variation better than changing one factor at a time??
• Which combination of roast factors gives the closest match to a target aroma pattern and TDS range??
• Does adding machine learning improve roast-level prediction compared with a simple threshold rule??
• How does run order affect measured roast quality when the same settings are repeated??

Basic Materials

• Home coffee roaster or school-access drum roaster with stable settings.
• Green coffee beans from one lot.
• Digital kitchen scale with 0.1 g accuracy.
• Coffee grinder with repeatable grind settings.
• Brewing setup with a pour-over cone or drip brewer.
• TDS meter or refractometer for coffee.
• Small gas sensor array or electronic nose kit.
• Thermometer or roast probe if the roaster does not log temperature.
• Airtight sample bags or jars for storing roasted beans.
• Spreadsheet software for recording batch conditions and results.

Advanced Materials

• Instrumented drum roaster with logged temperature and airflow control.
• Laboratory refractometer for coffee extraction measurements.
• GC-MS access for volatile compound profiling.
• Higher-resolution sensor array for aroma classification.
• Data logger for roast curves and environmental conditions.
• Reference roast samples for calibration and blind testing.
• Statistical software for DOE modeling and response surface analysis.
• Python environment for machine learning and cross-validation.
• Controlled storage containers with humidity monitoring.
• Standardized cupping materials for sensory verification.

Software & Tools

• Google Sheets: Organizes roast runs, tracks factor settings, and helps you spot patterns fast.
• Python: Fits DOE models, trains a roast-level classifier, and checks prediction error.
• JASP: Runs basic statistics, ANOVA, and simple regression without a steep learning curve.
• ImageJ: Analyzes color of roasted beans or brew color from consistent photos.
• Orange Data Mining: Builds beginner-friendly machine learning models with visual workflows.

Experiment Steps

1. Define the target roast profile you want to hit, and turn that goal into measurable outputs such as TDS, sensor pattern, and roast class.
2. Choose the few process factors you can control reliably, then set realistic high and low values for each one.
3. Plan a DOE layout that changes more than one factor at a time while keeping the total number of roasts manageable.
4. Build a measurement plan that keeps brewing, sensing, and sample storage consistent across every batch.
5. Create a calibration strategy so your sensor readings and roast labels connect to a real scale instead of loose judgments.
6. Decide how you will test repeatability, model error, and the gap between your target profile and the batches you actually make.

Common Pitfalls

• Changing bean age between runs, which adds variation that looks like a roast effect.
• Using inconsistent grind size or brew method, which makes TDS changes hard to trust.
• Letting the sensor array warm up differently each session, which shifts aroma readings.
• Mixing up roast color labels with actual measured roast level, which weakens the ML model.
• Running too few repeats, which hides whether a setting is truly repeatable or just lucky.

What Makes This Competitive

A stronger version of this project does more than compare roast settings. It tests interactions, like whether airflow matters more at one batch mass than another. It also checks model quality with proper validation, not just training accuracy. If you can show a repeatable process map that predicts roast outcome with fewer trials, the work starts to look like real process engineering.

Project Variations

• Swap the sensor array for a phone-based color analysis of bean surface color and compare how well color predicts roast level.
• Focus on one bean origin and test whether the DOE still finds a stable target profile across different roast dates.
• Replace the roast-level classifier with a sensory panel score and compare human ratings against TDS and sensor data.

Learn More

• MIT OpenCourseWare: Search for introductory courses in statistics, experimental design, and process improvement to learn DOE basics.
• NIST Engineering Statistics Handbook: Look up sections on design of experiments, calibration, and uncertainty analysis on the NIST site.
• USDA Coffee Research and Resources: Search USDA and university extension materials for coffee chemistry, roasting, and quality measurement.
• PubMed: Search review articles on coffee aroma compounds, roast chemistry, and sensory measurement.
• Introduction to Statistical Quality Control: Find a library or online textbook copy for control charts, process variation, and quality methods.