Onyx operates on the raw spectral data from each pixel in an EDS area map. The proprietary algorithm groups pixels into statistically distinct populations (Classes) based entirely on spectral similarity — not by comparison to a pre-built library, not by labels provided by a human, and not by a neural network trained on examples of “what a precipitate looks like.” This distinction matters. A CNN can only find what it has been trained to look for, requires a labeling step for every material system, and must be retrained whenever acquisition parameters change — accelerating voltage, magnification, working distance. For labs working across multiple alloys and instruments, this training debt often exceeds the cost of manual review itself. Onyx’s statistical approach eliminates it entirely. There is no labeling step, no per-system training, and no retraining. Because the model is statistical rather than supervised, it will identify every distinguishable phase in the dataset on the first encounter. This includes novel contaminants, unexpected intermetallics, and any phase not anticipated by the operator.

Onyx keeps the expert in the loop. Users inspect intermediate results and can act on beam interaction volume effects and topographic artifacts before quantifying phase area fractions, counts, and diameters. The model handles detection and measurement; interpretation stays with the domain expert.

The same engine runs on any alloy, any microscope, any dataset.