FTIR QA/QC with genAI: Application to the Minerals Industry
We examine the use of a class of generative models called 'variational autoencoders' to detect when a FTIR spectrum, classed as an unmineralized BIF from rock chips, may have been accidentally mislabelled.
Generative models have gained attention in recent years for their ability to create things like art, music, and realistic impersonations (e.g. faces).
This presentation explores how we can leverage that capability in order detect outliers or anomalies within geological datasets. We examine the use of a class of generative models called 'variational autoencoders' to detect when a FTIR spectrum, classed as an unmineralized BIF from rock chips, may have been accidentally mislabelled.
This type of work is important because it can be done prior (yes, it lives in the Bayesian world!) to being integrated into larger decision making models, whereby the outliers may invoke wrong predictions. Wrong predictions may result in significant consequences later on in the value chain and/or during the material processing stage.
We show how this method offered an initial screening of data in situations where it may not be possible for human intervention to keep up or analyse other spectra coming in rapidly.
