AEROSOL SCIENCE × MACHINE LEARNING × BIOSECURITY
AeroEvidence
From airborne signals to evidence for review.
An AI screening demonstrator that brings particle measurements, fluorescence and environmental context into one interpretable workflow.
SIMULATED SENSOR STREAM
Watch the signal evolve
Each frame represents one illustrative minute. Time and concentrations are demonstration values, not instrument observations.
SIZE-RESOLVED VIEW
Particle distribution
Illustrative optical size bins. These are not SMPS measurements or a calibrated biological size signature.
INTERPRETABLE SCORE
Why this prediction?
Signed contributions to model log-odds, relative to training-standardised features. Association in a synthetic model does not establish cause.
VALIDATION EVIDENCE
Performance on validation data
Confusion matrix
Reference: synthetic biological-like event vs other synthetic frames.
| Reference / prediction | Review flag | No flag |
|---|---|---|
| Biological-like | ||
| Other |
The biosecurity interpretation
In this demo, an alert starts a review of sensor quality and environmental context. Confirmation would require appropriate independent evidence through established laboratory and organisational procedures.
A missed synthetic event and an unnecessary review are both counted. No score on this page estimates the probability of a biological attack.
FRAME REVIEW LOG
Inspect every observation
Latest 12 frames| Minute | Scenario | Particles/L | Fluorescent % | Score | Flag | Synthetic truth |
|---|
METHOD, EXPERTISE & LIMITATIONS
Built around aerosol measurement judgement.
Aerosol science
Particle number, size distributions and environmental conditions help interpret an airborne signal. The demo connects to Prem’s work in aerosol dispersion and exposure assessment; it does not reproduce his experimental datasets.
Machine learning
A seeded synthetic generator creates overlapping background, biological-like, dust and fluorescent-interferent examples. A stratified 60/20/20 split supports training, validation and a test set sealed until the threshold is locked.
Biological assurance
Optical proxies can support screening. Species identity, viability and health significance require separate evidence. Non-biological fluorescence and unrepresented conditions can confound predictions.
Technical implementation and reproducibility
Seed 73, 1,800 examples. Logistic regression: 800 iterations, full-batch gradient descent, learning rate 0.12, L2 = 0.01, intercept unpenalised. Feature means and standard deviations are fitted on training data only. Inputs: fluorescence fraction, coarse fraction and natural log of particle concentration relative to an illustrative baseline of 500 particles/L. Point estimates and Wilson 95% intervals describe synthetic performance only.
Concentrations and particle-size histograms are generated for illustration. Labels come from the selected synthetic scenario, not measurements. High humidity is an out-of-training-distribution stress scenario; its model scores are especially unreliable. The simulated sequence is not a field validation dataset.
After revealing test results, retuning by resetting is exploratory. The browser-side seal is a workflow aid, not a security boundary. No live sensors, cloud AI API, external data, local storage or operational deployment are configured. Exported reports include model parameters, data provenance, frame readings and evaluation metrics. Portfolio demonstration built with AI-assisted development.
Scientific context
The research below motivates the need to consider fluorescence interference. It does not validate this simulator or its chosen synthetic distributions.