climate-weather/ai produced the result/Environmental Science & Technology 2026 · v3
Machine learning points to filter loading as main driver of levoglucosan desorption temperature
Researchers linked the shape of instrument signals from a year of Arctic aerosol measurements to 18 environmental and instrumental factors. Two tree-based models, probed with feature-importance methods, singled out how much material sat on the filter.
spectrum · one line per step, placed by what the step does · bright lines used AI
A Machine Learning Approach to Understand Thermal Desorption Profiles of Levoglucosan from FIGAERO–CIMS
Environmental Science & Technology, 2026
doi:10.1021/acs.est.5c18488 · record aix-00289 v3 · checked 2026-11-05
- AI was for
- Property prediction, Attribution
- Model family
- Random forest, Gradient-boosted trees, Linear model
- Checked by
- Held-out
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Tiny particles floating in the air, known as aerosols, affect clouds and climate, and one way to study them is to work out how easily their ingredients evaporate. A common instrument collects particles on a filter, then slowly heats the filter and watches which molecules come off at which temperature. The temperature at which the signal peaks is treated as a clue to how volatile the substance is. The snag is that this peak temperature may also shift for reasons that have nothing to do with the chemistry, such as the state of the instrument or the conditions outside. Untangling those influences from a long run of real measurements is awkward, because many of them vary together.
The team worked with levoglucosan, a molecule released when wood and other plant matter burns, measured over a year at the Zeppelin Observatory on Svalbard in the Arctic. From 443 heating curves they took the peak temperature as the quantity to explain, and gathered 18 candidate explanations spanning the weather, the physical and chemical make-up of the aerosol, and the instrument's own settings, including the mass of material collected on the filter.
Where AI came in
Three statistical models were fitted to predict the peak temperature from the 18 parameters: an ordinary multiple linear regression, a random forest and an XGBoost regressor. The latter two are collections of decision trees, which can capture bends and interactions that a straight-line fit cannot. Models were trained on 80 per cent of the data and judged on the remaining 20 per cent. The random forest explained 0.65 of the variation (with a typical error of 17.24 °C) and XGBoost 0.66 (17.22 °C), against 0.4 for the linear regression.
The models were then interrogated rather than simply used to predict. Two standard techniques, permutation feature importance, which scrambles one parameter at a time to see how much accuracy drops, and SHAP values, which apportion each prediction among the inputs, were applied to the fitted trees. Both ranked the mass loading on the filter highest, with lower loadings going with higher peak temperatures. The linear model, used as a comparison, instead put temperature-related parameters first. The authors recommend controlling the mass collected, suggesting around 150 ng, when peak temperature is used to infer volatility.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
Three regression models (multiple linear regression, a random forest regressor and an XGBoost regressor) were applied to 443 levoglucosan thermograms from a year of FIGAERO–CIMS measurements at the Zeppelin Observatory on Svalbard, relating the temperature of maximum signal (T max) to 18 meteorological, aerosol physical, aerosol chemical and instrumental parameters. On the test data the random forest reached a coefficient of determination of 0.65 (RMSE 17.24 °C) and XGBoost 0.66 (RMSE 17.22 °C), against 0.4 for the linear regression. Permutation feature importance and SHAP values from both machine learning models ranked the mass loading on the FIGAERO filter as the most important parameter for T max variation, with lower mass loadings associated with higher T max, while the linear model instead ranked temperature-related parameters highest. The authors recommend controlling the mass collected on the filter, suggesting loadings of around 150 ng, when T max is used to derive volatility.
How AI was used
Thermograms of the ion I(C6H10O5)− and 18 environmental and instrumental parameters were homogenised onto the 2.5 h FIGAERO sampling intervals, and T max was extracted from each thermogram after smoothing with a Savitzky-Golay filter, with a per-sample weight derived from the deviation between the original and smoothed curve. T max was then used as the target variable for three fitted models: an ordinary-least-squares multiple linear regression on MinMaxScaler-normalised inputs, and a random forest regressor and an XGBoost regressor trained on absolute parameter values with the noise-derived sample weights. For the two tree-based models the data were split 80/20 into training and test sets, and hyperparameters were selected by randomized search followed by grid search, giving 100 trees with maximum depth 20 for the random forest and 200 trees with maximum depth 5 and learning rate 0.025 for XGBoost. Performance was measured by the coefficient of determination and RMSE. Parameter relevance was then obtained by running permutation feature importance and SHAP on the trained regressors over the held-out data, and the resulting importances were min-max normalised so they could be compared against each other and against the linear model's regression coefficients.
The shape of the work
Structural · the record, drawn
Assemble thermograms and ancillary parameters
Obtaining raw data, whether by measurement, download or retrieval.
In total, 18 parameters (Table S1 in the Supporting Information) were part of the analysiswhere the paper describes this · verbatim
Derive T max target and align parameters in time
Cleaning, filtering, normalising or labelling data already obtained.
The smoothed version was obtained by applying a Savitzky-Golay filter, a method to smooth noisy datawhere the paper describes this · verbatim
Fit regression models to T max
Fitting model parameters, including fine-tuning an existing model. The AI stood in for statistical model.
the best hyperparameters were obtained by using RandomizedSearchCV and GridSearchCV from the scikit-learn library (version 1.7.2)where the paper describes this · verbatim
Evaluate predictions on held-out data
Testing outputs against ground truth.
For both regressors, the data were divided into 80% for training and 20% for testingwhere the paper describes this · verbatim
Attribute T max variation to parameters
Extracting understanding from model behaviour. The AI stood in for statistical model.
the permutation feature importance (PFI), and the Shapley additive explanations (SHAP) values were usedwhere the paper describes this · verbatim
Compare normalised importances across models
Extracting understanding from model behaviour.
the importances of the parameters per model were normalized using the MinMaxScaler from the scikit-learn librarywhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The paper's central claim, that mass loading on the FIGAERO filter is the most relevant parameter for levoglucosan T max, is produced by the permutation feature importance and SHAP analyses of the random forest and XGBoost regressors; the linear model is run as a comparator and disagrees.
three different modeling approaches were used: a multiple linear regression (MLR) model and two machine learning modelswhere the paper describes this · verbatim
For both regressors, the data were divided into 80% for training and 20% for testingwhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- ComputeThe hardware or time used is not stated.
- How many were testedThe paper gives no count of what was tested.
- What step 4 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00289, version 3, checked by a person on 2026-11-05. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error
Cite this record: AIxSci (2026). Machine learning points to filter loading as main driver of levoglucosan desorption temperature. Record aix-00289, version 3, checked 2026-11-05. https://aixsci.org/articles/aix-00289/