A Study on an Explainable Causal-Enhanced LLM Agent for Predicting the Forming Quality of Automotive Component Materials

Abstract

Raw material composition, heat treatment, cooling, equipment parameters and die condition have an impact on the forming quality of materials for automotive components. Current prediction models can detect the defect or deviation in the performance, but are not so much supported for finding out the root cause of anomaly and for making the changes in the process. In this paper, a causal enhanced LLM Agent explainable analysis framework is proposed. The study builds a multi-source manufacturing dataset that includes the composition of raw materials, heat treatment curves, forming pressure, cooling time, die condition, inspection outcome, and rework records;Random Forest, Multi-Layer Perceptron, XGBoost, and TabNet are used to predict strength deviation, surface defect grades, and batch pass rates;SHAP, NOTEARS, and counterfactual reasoning are combined to identify the key process variables and their influence paths;In addition, a RAG-LLM Agent is built to fuse the process specifications, material handbooks, historical anomaly cases, and model interpretation results, which helps to generate quality root cause analysis and process adjustment recommendation. As an example, a material forming manufacturing line for automotive parts had 68,000 batch records, 42 process variables and 11 quality metric categories included in the experiment. Across ten repeated group-stratified evaluations, XGBoost achieved a macro-F1 score of 0.892 with a 95% confidence interval of 0.888–0.896, compared with 0.838 for Random Forest, 0.856 for MLP, and 0.874 for TabNet. The increase over Random Forest was 0.054 in absolute terms and 6.4% in relative terms, and the paired bootstrap test confirmed that the difference was statistically significant at p<0.001.The causal constraint also improved the key variable identification consistency by 18.7%, Among 200 independently reviewed reports, the Agent’s primary root-cause conclusion matched the engineer consensus in 169 cases, yielding an exact agreement rate of 84.5% and a Cohen’s kappa of 0.781. The time taken to identify problems was also reduced from 47 minutes to 19 minutes. The study shows that output of causal reasoning and the LLM Agent can improve interpretability of the prediction of material quality and the impact of process interventions.