ParticleTensorAI® uses artificial intelligence to analyse multidimensional representations of time-resolved light-scattering signals. The Particle Tensor organises the measurement data; the AI model learns patterns in those data and relates them to a defined classification or estimation task.
From Particle Tensor to Model Output
An AI model receives Particle Tensors as input. During development, the model is provided with measurement data and suitable reference information. It learns statistical relationships between patterns in the tensors and the corresponding reference labels or values.
Once trained and validated, the model can analyse new Particle Tensors and produce an output defined for the measurement task. Depending on the model, this may be a classification, such as assigning a measurement to a known condition, or an estimate of a target property.
Training with Reference Data
In supervised learning, each training example is paired with reference information. This information tells the model what the Particle Tensor represents—for example, a known material condition or a reference measurement result.
The model adjusts its internal parameters during training to reduce the difference between its predictions and the reference information. The learned relationship is then assessed using data that were not used to train the model.
Convolutional Neural Networks as One Example
A published ParticleTensorAI® study used convolutional neural networks (CNNs) to analyse image-based Particle Tensor representations. In that study, detector signals were mapped to RGB image channels. The tensors were labelled with average coal-particle size and mass concentration for a coal monoethylene glycol slurry spray.
The CNNs were trained for classification tasks. Validation experiments reported prediction accuracies of up to 95% under the conditions investigated in that study. This is a result for that specific research application and does not represent a general accuracy specification for other materials, measurement setups, or models.
The Role of the AI Model
The AI model does not replace the optical measurement or create information that is absent from the recorded signals. It analyses patterns in the measured data and relates them to the reference information used during its development.
The tensor format, detector signals, reference labels, and model architecture together define what the model can learn. A model trained to classify one set of conditions should not be assumed to provide validated results for a different task without suitable testing.
Validation and Limits
Validation evaluates how well a trained model performs on data that were kept separate from the training process. The validation data should represent the materials and measurement conditions relevant to the intended application.
- Training data determine which signal patterns the model can learn.
- Reference information defines the output the model is trained to produce.
- Validation data are used to assess performance beyond the training examples.
- Changes in material, process conditions, or measurement setup may require additional validation.
Summary
In ParticleTensorAI®, AI analysis follows the creation of the Particle Tensor. The tensor organises time-resolved signal information, while the trained model identifies relationships between tensor patterns and reference data. The resulting output is meaningful within the task and conditions for which the model has been developed and validated.
Further Reading
For a published example of CNN-based analysis of Particle Tensor representations, see “AI-assisted monitoring of coal particles within droplets in a coal monoethylene glycol slurry spray” in the KITopen repository.