⚡ electrical engineering

Forecasting Lab Climates: A Machine Learning Decision Support System for Humidity-Sensitive Experiments

This paper presents a site-specific machine learning framework that forecasts indoor laboratory relative humidity up to seven days in advance by leveraging local weather data and calendar-based occupancy patterns, thereby enabling researchers to better plan humidity-sensitive electrospinning experiments and improve reproducibility.

Original authors: Lukas Tilch, Maximilian Heumann, Jan Fabian Drexler, Marc Mueller

Published 2026-07-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Lukas Tilch, Maximilian Heumann, Jan Fabian Drexler, Marc Mueller

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a scientist trying to spin tiny, invisible threads of plastic (called electrospinning) in your lab. The secret to making perfect threads isn't just your fancy machine; it's the air around you. If the air is too dry or too humid, your threads might get bumpy, break, or turn into a weird film instead of fibers. It's like trying to bake a cake where the humidity in the kitchen changes the recipe while you're mixing the batter.

The problem? Scientists usually only check the weather after their experiment is done, realizing too late that a rainy Tuesday ruined their perfect fibers. They are flying blind, hoping the air stays stable.

Enter the "Crystal Ball for Labs." This paper introduces a smart computer system that acts like a weather forecaster, but instead of predicting rain for your picnic, it predicts the humidity inside your specific laboratory room up to 7 days in advance.

The Big Discovery: One Size Does Not Fit All

The researchers set up sensors in four different labs inside the same building. You might think, "Hey, they're in the same building, so the air should be the same, right?" The paper says nope.

Think of the building like a giant house with four different rooms. Even though the central air conditioner is the same, one room faces north, another faces east, one is small and cozy, and another is huge and open. The paper argues that you cannot use one single "magic formula" for all of them. Just like a north-facing bedroom stays cooler than a south-facing living room, each lab reacts to the outside weather in its own unique way. So, the team trained four separate computer brains, one for each room, to learn their specific personalities.

How the Crystal Ball Works

The system doesn't just guess; it looks at the outside world to see what's coming. It watches the outdoor dew point (how much moisture is in the air outside), the sun's angle, and even the calendar.

Here is the cool part: The computer noticed that humans are part of the weather too. When people are in the lab (working days), the temperature and humidity behave differently than when the lab is empty (weekends). The system uses a "calendar trick" to guess if people are around. It found that on busy days, the air gets more "jumpy" (more variable) because of people opening doors and running experiments. By looking at the day of the week, the computer can anticipate these human-induced bumps in the air.

How Good is the Prediction?

The paper tested this system with a "time-travel" challenge: Can the computer predict the future based on past data?

Short-term (Today to 2 days): The system is a superstar. For the Random Forest model (one of the smart algorithms they used), the prediction was incredibly sharp, with an error of only 0.17 percentage points of humidity on day 0. It's like guessing the temperature with almost perfect accuracy.
Medium-term (3 to 4 days): The system is still very good, but the crystal ball gets a little foggy. The error creeps up to about 2.86 percentage points by day 4.
Long-term (7 days): The fog gets thicker. By day 7, the error is around 4.93 percentage points. The system admits, "I'm not 100% sure what the air will be like a week from now."

The paper also compared different types of computer brains. A simple "linear" model (like a straight line) failed miserably, getting worse and worse until it was basically guessing randomly. But the fancy, non-linear models (like XGBoost and Random Forest) handled the complex, wiggly nature of the air much better.

The "Warning Light" Dashboard

The researchers didn't just build a math model; they built a dashboard for the scientists. Imagine a graph showing the next 7 days.

A magenta line shows the predicted humidity.
Shaded bands around the line show how unsure the computer is. On day 1, the band is thin (high confidence). By day 7, the band is wide (low confidence).
Dashed lines show the "safe zone" for your specific experiment.

This lets a scientist look at the screen and say, "Oh, next Tuesday the humidity might jump out of the safe zone. I should probably do my experiment on Monday instead, or change my recipe."

What the Paper Says It Is NOT

It's important to know what this system doesn't do.

It doesn't fix the air: The system doesn't automatically turn on the AC or dehumidifier. It's a "decision support" tool, meaning it gives you a heads-up so you can decide what to do.
It's not a universal magic wand: The specific computer brains trained in these four labs won't work perfectly in a different building. The paper explicitly states that if you want to use this in your own lab, you have to train a new model with your local data. The "recipe" is portable, but the "cake" is specific to your kitchen.
It doesn't know your specific plastic: The system predicts the humidity, but it doesn't tell you exactly how that humidity will change your specific fiber. That part is still up to the scientist's expertise.

The Bottom Line

The paper suggests that we can't just ignore the weather when doing delicate science. Even inside a building, the outside air sneaks in and changes things. By using a smart, room-specific forecast that accounts for the calendar and the sun, scientists can stop being surprised by their experiments. It's not a perfect crystal ball that sees the future with 100% certainty, but it's a powerful tool that turns "hoping for the best" into "planning for the likely."

The authors note that while the results look promising, they only have data from one year. They suspect the system might need to be re-tuned if the seasons change drastically or if the building's heating and cooling systems act differently in winter versus summer. But for now, it's a solid step toward making science more reliable, one humidity forecast at a time.

Technical Summary: Forecasting Lab Climates for Humidity-Sensitive Experiments

Problem Statement

The reproducibility of electrospinning, a process for producing micro- and nanofibrous materials, is highly sensitive to ambient relative humidity (RH). RH influences critical process parameters such as solvent evaporation, jet stability, and fiber solidification, ultimately determining fiber morphology, mechanical behavior, and crystalline structure. While Heating, Ventilation, and Air Conditioning (HVAC) systems are intended to buffer indoor environments from outdoor weather, indoor humidity in laboratory settings often retains a measurable dependence on external conditions over hourly to seasonal timescales.

Currently, researchers typically document environmental variables retrospectively. This reactive approach limits the ability to anticipate weather-driven shifts in laboratory conditions, potentially causing experiments to drift outside morphology-sensitive operating ranges before operators recognize the risk from current sensor readings alone. There is a lack of frameworks that combine indoor-climate predictability with electrospinning requirements to support proactive laboratory decision-making.

Methodology

The study proposes a machine learning-based Decision Support System (DSS) designed to forecast indoor laboratory RH using local meteorological measurements, forecasts, and internal climate data.

Data Acquisition and Preprocessing

Setting: Four laboratories (A–D) within a single centrally conditioned building at Leibniz University Hannover were selected. These rooms varied in orientation (north/east), size, and usage patterns.
Sensors: Indoor climate data (temperature, RH, pressure) were collected at 1-minute resolution using Bosch BME280 sensors.
Meteorological Inputs: Outdoor data were sourced from the Deutscher Wetterdienst (DWD) for training (10-minute resolution) and Open-Meteo for operational forecasting (hourly resolution).
Preprocessing: Data were cleaned, synchronized, and resampled to 1-minute resolution. Gaps of 20 minutes or longer were flagged and excluded.

Feature Engineering

The model utilized a set of exogenous predictors without relying on lagged indoor humidity terms (to ensure portability to labs with limited monitoring infrastructure):

Moisture-State Variables: Dew-point temperatures calculated via the Magnus-Tetens approximation.
Temporal Descriptors: Circular encodings for hour of day and day of week to capture cyclical patterns.
Solar-Exposure Variables: Effective solar impact ( $I_{eff}$ ) calculated based on solar zenith angle, azimuth, and window orientation (vertical).
Weather Memory: 6-hour rolling averages and 1-hour first-difference features for outdoor temperature and RH to capture recent history and rate of change.
Occupancy Proxies: Calendar-based indicators for working days, public holidays, and university closures to approximate human activity patterns.

Model Development and Validation

Algorithms: The study compared Multilinear Regression (LR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost).
Validation Strategy: A blocked temporal validation was employed. For each calendar month, all samples were withheld as an external test block while the model was trained on the remaining months. This prevented data leakage and tested the model's ability to generalize across seasonal regimes.
Uncertainty Quantification: Prediction intervals (50%, 90%, 95%, 99%) were constructed empirically using out-of-sample forecast errors from archived past forecasts, calibrated separately for each forecast lead day (0–7 days).
Occupancy Analysis: A Wasserstein-distance heatmap was used to analyze temperature distributions between working and non-working days, serving as a proxy for human-induced climate variability.

Key Results

Humidity Drift and Variability

Seasonal Variation: Daytime RH swings (difference between max and min RH from 08:00 to 18:00) varied significantly by season and room. Winter swings were smallest (median ~2–5%), while summer swings were largest (Lab A reached medians of 14–15%, with individual days exceeding 30%).
Room Specificity: Significant differences existed between laboratories. Lab A (north-facing, standard use) showed the largest summer swings, while Lab D (east-facing) showed higher winter medians.
Occupancy Impact: Temperature analysis revealed that working days exhibited broader distributions and systematic deviations from non-working baselines, particularly during core daytime hours. This confirmed that human activity introduces learnable signals into the indoor climate.

Model Performance

Accuracy: The XGBoost model trained on Lab C data achieved high accuracy, with $R^2$ values between 0.9179 and 0.9927 and Mean Absolute Errors (MAE) between 0.532 and 2.154 percentage points RH.
Seasonal Performance: Performance was strongest in winter and late autumn (MAE < 0.8) and weaker in late spring and summer (MAE up to 2.154 in September).
Forecast Horizons:
- Short-term (Day 0–2): Random Forest outperformed XGBoost, achieving the lowest MAE (0.17) and narrowest prediction intervals.
- Long-term (Day 3–7): XGBoost generally outperformed RF, showing greater robustness as the forecast horizon extended.
- Baseline: Multilinear Regression performed poorly, with $R^2$ dropping to negative values at longer horizons, indicating that indoor humidity dynamics are non-linear.
Feature Importance: Outdoor dew-point variables (specifically 1h, 2h, and 3h moving averages) were the most influential predictors, followed by outdoor temperature and solar geometry.

Decision Support Output

The system generates a 7-day forecast visualization for operators. It displays the predicted indoor RH trajectory alongside the outdoor RH signal and nested, lead-day-specific prediction intervals. This allows researchers to identify potential RH excursions, assess the timing and confidence of these shifts, and plan experiments accordingly.

Significance and Claims

The paper positions its contribution not as a universally transferable model, but as a portable forecasting and decision-support workflow.

Site-Specific Modeling: The study demonstrates that indoor-outdoor transfer functions are not uniform across a building. Separate models for each room are necessary to account for local boundary conditions (orientation, ventilation, thermal buffering).
Proactive Planning: The DSS shifts the paradigm from reactive monitoring to proactive planning. By providing lead-time-specific forecasts with uncertainty bounds, it allows researchers to anticipate weather-driven RH shifts that could compromise electrospinning quality before they occur.
Exogenous Reliance: The framework successfully reconstructs indoor RH using only exogenous features (outdoor weather, calendar, solar geometry) without lagged indoor history. This design choice lowers the barrier to adoption for laboratories lacking complex historical data or continuous indoor monitoring infrastructure.
Operational Utility: The system is designed to support human decision-making rather than autonomous control. It provides the necessary context for researchers to adjust process parameters (e.g., polymer concentration, solvent composition) or reschedule experiments based on predicted environmental stress.

The authors acknowledge limitations, including the single-year dataset, the use of calendar proxies for occupancy, and the lack of polymer-specific threshold integration. They conclude that while the trained models are site-specific, the underlying workflow—data acquisition, feature engineering, uncertainty calibration, and visualization—can be adapted to other laboratory settings using local data.