climate-weather/ai in a supporting role/Environmental Science & Technology 2025 · v2
Bacteria-derived aerosol matter made to produce oxidants, then modelled from its optical fingerprints
Researchers irradiated watery extracts of organic matter from bacteria cultured from Hong Kong roadside particles and measured the reactive oxidants formed. Machine learning models were then fitted to link the samples' optical and chemical properties to that oxidant production.
spectrum · one line per step, placed by what the step does · bright lines used AI
Efficient Production of Reactive Oxidants by Atmospheric Bacterial-Derived Organic Matter in the Aqueous Phase
Environmental Science & Technology, 2025
doi:10.1021/acs.est.5c01526 · record aix-00281 v2 · checked 2026-11-05
- AI was for
- Property prediction
- Model family
- Gradient-boosted trees, Random forest, Linear model
- Checked by
- Held-out
- Code
- not reported
AI processed or interpreted data, but the main finding does not rest on it.
What this research was about
Tiny airborne particles, including the fraction known as PM2.5 (particles smaller than 2.5 micrometres across), carry water, chemicals and living bacteria. Inside cloud droplets and the thin films of water on such particles, sunlight can turn dissolved organic material into short-lived reactive species: hydroxyl radicals, singlet oxygen and so-called triplet excited states, which are molecules holding absorbed light energy for a moment before passing it on. These oxidants drive much of the chemistry that ages aerosol. Working out which components of a soupy natural mixture are responsible is hard, because thousands of unidentified molecules absorb light together.
The team looked at two kinds of bacterial material: organic matter from inside the cells and the sticky extracellular polymeric substances bacteria secrete around themselves. They cultured bacteria from PM2.5 collected at a Hong Kong roadside in summer and winter, along with pure strains, extracted both materials into water and split them by molecular weight. They then shone light on the solutions with chemical probe compounds present, substances that react with one oxidant in a known way so its amount can be inferred, and measured how much of each oxidant formed. Alongside this they recorded absorption and fluorescence spectra.
Where AI came in
The oxidant measurements themselves involved no AI. The modelling came afterwards. The researchers took the optical and chemical descriptors of each sample, including absorbance and fluorescence indices, the contributions of six fluorescence components and the protein and polysaccharide content, and used them as inputs to predict the measured oxidant yields and steady-state levels. They fitted both multiple linear regression, a standard statistical fit, and five tree-based machine learning methods, which build predictions by stacking many simple yes/no splits on the input values. Categorical boosting scored best in cross-validation and was tuned further; eighty per cent of the data trained the models and twenty per cent tested them, giving test R values of 0.76 to 0.90.
The machine learning stood in for a purely linear statistical model, allowing curved and interacting relationships between sample properties and oxidant production. Because such models are opaque, the team examined them with feature importance rankings, SHAP values (a way of apportioning a prediction among its inputs) and partial dependence plots, which trace how a prediction shifts as one input varies. These pointed to measures linked to aromatic ring content and to molecular size as the properties associated with oxidant formation.
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
Aqueous extracts of cellular organic matter (COM) and extracellular polymeric substances (EPS) from bacteria cultured from Hong Kong PM2.5 and from pure cultures were irradiated with chemical probes to measure production of hydroxyl radicals, singlet oxygen and organic triplet excited states. Average singlet oxygen apparent quantum yields were (4.77 ± 0.58)% for bulk COM and (4.12 ± 0.73)% for bulk EPS, reaching (10.78 ± 0.82)% and (8.89 ± 0.88)% for the molecular-weight-fractionated COM and EPS samples, and quantum yields and steady-state concentrations decreased with increasing molecular weight. Multiple linear regression and tree-based machine learning models were then fitted to relate the measured optical and chemical parameters to oxidant production; the tuned categorical boosting models reached test-set R values of 0.76 to 0.90, and feature importance, SHAP and partial dependence analyses pointed to parameters related to aromaticity (FIX, SUVA254) and molecular size (E2/E3).
How AI was used
After the photochemistry and spectroscopy measurements were complete, the measured optical and chemical parameters of the bulk and molecular-weight-fractionated COM and EPS samples (absorbance indices (E2/E3, SUVA254), fluorescence indices (BIX, FIX, HIX), PARAFAC fluorescence component contributions, and protein and polysaccharide concentrations) were used as features to model the log-transformed apparent quantum yields and steady-state concentrations of the three oxidants. Multiple linear regression fits were developed in SIMCA 14.1 and Prism 8, with predictors chosen by orthogonal partial least-squares analysis. In parallel, five tree-based algorithms (random forest, gradient boosting decision tree, light gradient boosting machine, extreme gradient boost and categorical boosting) were implemented in Python 3.11 on top of scikit-learn (v.1.1.3), with 80% of the data used for training and 20% for testing; categorical boosting was selected on cross-validation scores obtained with default hyperparameters, and its hyperparameters were then tuned by Bayesian optimization. The resulting models were examined with tree-based feature importance analysis, SHAP and partial dependence plots to identify which optical and chemical parameters influenced the predictions and whether the relationships were linear or nonlinear. Spearman rank correlation analysis in Origin and Student's t-tests in Prism 8 were run alongside the models as conventional statistics.
The shape of the work
Structural · the record, drawn
Collect PM2.5 and culture bacteria
Physical execution, by hand or by robot.
PM2.5 samples were collected for 24 h at a flow rate of 85 m/hwhere the paper describes this · verbatim
Extract and size-fractionate COM and EPS
Cleaning, filtering, normalising or labelling data already obtained.
the filtered COM and EPS extracts were ultrafiltered using five centrifugal size filterswhere the paper describes this · verbatim
Measure optical and chemical properties
Obtaining raw data, whether by measurement, download or retrieval.
UV–visible absorption spectra (200–800 nm with 1 nm increments) were measured by a UV–visible absorption spectrophotometerwhere the paper describes this · verbatim
Decompose fluorescence spectra with PARAFAC
Encoding data into features, descriptors, embeddings or graphs.
Parallel factor (PARAFAC) analysis was performed on the EEM fluorescence spectra to determine the composition of the fluorophores in the COM and EPS samples.where the paper describes this · verbatim
Quantify oxidant production by probe photochemistry
Physical execution, by hand or by robot.
Benzene (1–3 mM) was used as the chemical probe for •OH, and the formation of phenol was measured during the experiments.where the paper describes this · verbatim
Fit multiple linear regression models
Fitting model parameters, including fine-tuning an existing model.
Multiple linear regression (MLR) models for ΦRI and [RI]ss were developed in SIMCA 14.1 and Prism 8 by using measured optical and chemical parameterswhere the paper describes this · verbatim
Select, train and tune tree-based ML models
Fitting model parameters, including fine-tuning an existing model. The AI stood in for statistical model.
Since the CB algorithm had substantially higher cross-validation scores than the other ML algorithmswhere the paper describes this · verbatim
Interpret models with feature importance, SHAP and PDP
Extracting understanding from model behaviour.
Details of the tree-based feature importance analysis, Shapley’s additive interpretation (SHAP), and partial dependence plots (PDP) used to interpret the resultswhere 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.
Oxidant quantum yields and steady-state concentrations were obtained by chemical-probe photochemistry experiments; the machine learning and regression models were applied afterwards to relate and predict those measured quantities, so the central measurements do not depend on the models.
five commonly used candidate tree-based machine learning (ML) algorithmswhere the paper describes this · verbatim
80% of the data was used for training, while the remaining 20% was used for testing.where 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.
- Version of Categorical boosting (CatBoost, CB)Which version of the model was used is not stated.
- Version of Random forest (RF)Which version of the model was used is not stated.
- Version of Gradient boosting decision tree (GBDT)Which version of the model was used is not stated.
- Version of Light gradient boosting machine (LGB)Which version of the model was used is not stated.
- Version of Extreme gradient boost (XGB)Which version of the model was used is not stated.
- Version of Multiple linear regression (MLR)Which version of the model was used is not stated.
- What step 6 replacedThe paper gives no basis for what the AI stood in for.
- What step 8 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00281, version 2, 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). Bacteria-derived aerosol matter made to produce oxidants, then modelled from its optical fingerprints. Record aix-00281, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00281/