ParticleTensorAI® is based on the idea that time-resolved light-scattering signals contain information beyond the properties of a single detected particle event. By retaining sections of the recorded signals and arranging them as multidimensional data structures, an AI model can analyse patterns across time and, where available, across detector channels.
1. Light Scattering Creates Time-Dependent Signals
When a droplet or particle passes through a defined light field, it scatters light. The intensity recorded by an optical detector changes as the particle moves through the illuminated region. The resulting signal therefore varies over time.
The shape and intensity of the signal are influenced by the interaction between the particle and the light field. Depending on the measurement task, this signal can contain information related to particle or droplet properties, material characteristics, and the conditions of the particle flow.
When several detectors observe scattered light from different directions, they produce separate signal streams. These streams can provide complementary information about the same measurement interval.
2. From Individual Events to Signal Sequences
In event-based measurement, a signal is examined to identify individual particle interactions. The system can then calculate selected parameters from each detected event.
ParticleTensorAI® uses a different evaluation concept. A section of the time-dependent signal can be retained for analysis without first requiring every particle interaction to be identified and separated as an individual event. This is referred to as trigger-free signal-sequence analysis.
Trigger-free analysis does not mean that particles do not generate separate physical interactions. It means that the AI evaluation can operate on the recorded signal sequence without requiring a trigger-based event separation step as its starting point.
3. Representing Signals as a Particle Tensor
Before AI analysis, the recorded data are arranged into a multidimensional representation called a Particle Tensor. The representation preserves selected information from the signal sequence, such as signal intensity, temporal structure, and relationships between detector channels.
The tensor structure depends on the measurement and analysis task. Its dimensions and data mapping are defined so that the relevant signal information can be presented consistently to the selected AI model. A tensor is therefore a structured representation of measured data; it is not, by itself, a physical measurement result.
4. AI-Based Evaluation
An AI model analyses the Particle Tensor to classify a condition or estimate a target property. The model learns relationships between the tensor patterns and reference information supplied during its development.
For a supervised model, the training data are paired with known labels or reference values. The trained model can then evaluate new Particle Tensors that fall within the validated measurement domain. The reliability of its output depends on the quality and coverage of the training data, the measurement conditions, and the validation procedure.
5. Relationship Between Signal Physics and Model Output
The AI model does not replace the optical measurement principle. It evaluates patterns in signals produced by light scattering. The physical measurement defines what information can be recorded; the tensor representation organises that information; and the AI model relates the recorded patterns to the selected analysis objective.
For this reason, the output must be interpreted in relation to the specific measurement setup, material, process conditions, reference data, and validation results. A model developed for one application should not automatically be assumed to provide validated results for a different application.
Research Example
A published study investigated coal particles inside droplets in a coal–monoethylene glycol slurry spray. Detector signals were transformed into Particle Tensor representations and used to train convolutional neural networks for classification related to average coal-particle size and mass concentration.
The study reported prediction accuracies of up to 95% in validation experiments under the tested conditions. This is a result from that specific proof-of-principle investigation, not a general performance value for every ParticleTensorAI® application.
Read the research paper in the KITopen repository.
Summary
- Moving droplets and particles generate time-dependent light-scattering signals.
- ParticleTensorAI® retains signal sequences for analysis without requiring every event to be separated first.
- The signals are arranged into a multidimensional Particle Tensor.
- An AI model analyses the tensor in relation to a defined measurement objective.
- Model outputs require suitable reference data and validation for the relevant measurement conditions.