Data2Value: Explainable, Inline Image–Based Monitoring for Robust Batch Crystallization Keywords: Crystallization; Inline image analysis; Multimodal deep learning; XAI; Digital twin; Batch-to-batch learning Crystallization not only governs product purity, but also particle size and shape. The latter two properties influence filterability, downstream handling and, for APIs, the sensitive solid form with regard to regulatory and intellectual property issues. Yet batch-to-batch variability persists because critical “crystallization parameters” (onset/induction, nucleation/growth dynamics, crystal size/shape/morph fraction) are rarely measured inline; conventional process analysis technology (e.g., turbidity, focused beam reflectance measurement) provides only partial coverage. Data2Value closes this gap by combining inline, image-based particle analytics with a multimodal, explainable deep-learning classifier that delivers early warnings and root-cause indications in real time. The classifier fuses time-series features (image-derived size/shape trajectories, process signals) with static context (e.g., supplier, recipe, operator) and provides human-interpretable attributions via explainable AI. To mitigate the problem of the limit availability of plant data, the multimodal deep learning classification model (see Fig. 1) are pre-trained on synthetic trajectories generated by a first-principles digital twin (population balance equation-based), then fine-tuned on laboratory and mini-plant batches with continuous acquisition; models are updated batch-to-batch to track parametric drift and plant-model mismatch. We outline the laboratory validation of the image pipeline, the curated reference database, and the industrial mini-plant integration, and we present planned KPIs for impact (false-batch reduction, cycle-time, rework/energy). The approach is sensor-agnostic and transferable to other particulate unit operations (e.g., precipitation, granulation, spray drying). This poster outlines the Data2Value project, detailing its architecture, data flows and validation strategy. It also highlights the pathway from in-line visibility to operational and explainable decisions in the batch crystallization process.
Ferdinand Breit (Fri,) studied this question.