LLM‑Style Issue Prediction Using JMP Neural Modeling (2026-US-30MP-2970)

Manufacturing-fleet issue resolution often relies on manual interpretation of unstructured engineering logs, where outcomes are influenced by inconsistent descriptions or subjective categorization. This study demonstrates how neural predictive modeling in JMP 19 can be applied as an LLM‑style inference framework to improve issue prediction consistency.

Rather than generating free‑text responses, this approach focuses on predicting likely solutions using structured text features derived from historical engineering records. We begin with Text Explorer to tokenize issue descriptions and extract keywords and term frequencies – functionally analogous to the tokenization of large language models. These extracted features are then used as predictor variables in a neural predictive model, with the known issue solution defined as the response.

Given the limited nature of historical data, K‑fold cross validation is employed to obtain more reliable estimates of model performance and generalization. Initial attempts render under‑ or mis‑tuned models exhibiting high misclassification rates and weak generalized R². Through iterative optimization, the final model demonstrates low and consistent misclassification, indicating stable generalization and meaningful separation among outcome classes.

While not a true large language model, the result exhibits LLM‑style behavior by learning relationships between issues and historical outcomes. The model captures insights into the effects of chamber type and variability on resolution paths. This work illustrates how JMP can be used to reinforce engineering knowledge, accelerate root‑cause classification, and support more consistent decision making for Hybrid Bonder fleet support – without the complexity of full‑scale LLMs.