Data2Value: Explainable, Inline Image-Based Monitoring for Robust Batch Crystallization

Poster presented at ISIC 2026 (International Symposium on Industrial Crystallization), Budapest, Hungary. Crystallization is a key unit operation for the isolation and purification of active pharmaceutical ingredients, with more than 80 % of all APIs manufactured via crystallization. Beyond purity, crystal size and shape govern bioavailability, flowability, agglomeration tendency and filterability, and they dictate downstream filtration and drying. Industrial practice nevertheless shows pronounced batch-to-batch variation despite formally identical process operation, causing off-spec batches, rework, increased raw material and energy demand and reduced yield. The suspected root causes are hardly accessible disturbance variables that remain invisible to standard inline sensors. The Data2Value project addresses this gap along four lines. First, a systematically validated inline image-based measurement system for crystal size and shape in a lab-scale stirred crystallizer with optical access, time-synchronized with turbidity, pH, temperature, torque and stirrer speed. Second, a first-principles digital twin of the stirred crystallizer built around a two-dimensional population balance in size and shape, covering primary and secondary nucleation, anisotropic growth, agglomeration, breakage and polymorph transformation, used as a source of synthetic training data. Third, a multimodal deep learning classification model that fuses the image-based process analytical technology signal, process time series and context data into a fault diagnosis. Fourth, explainability by design, so that each alarm is attributed to the responsible features and time windows and yields a cause hypothesis rather than a black-box score. Expected outcomes are a validated inline image-based PAT benchmarked against established offline reference methods, an explainable early warning system that detects off-spec batches while they are still running and attributes them to their root cause, and an open, curated and annotated crystallization dataset together with a modular, sensor-agnostic toolbox transferable to other particulate processes.