LSTM-based early warning system for ceramic firing defects: a time-series approach

Area of Science:

• Materials Science
• Ceramic Engineering
• Educational Technology

Background:

• Ceramic firing is a key materials science skill but has long cycle times, delaying student feedback.
• High defect rates (~40%) in student ceramic firings exceed industrial standards.
• Current pedagogical methods rely on retrospective analysis, hindering timely intervention.

Purpose of the Study:

• To develop an early warning system for ceramic firing defects using machine learning.
• To improve pedagogical strategies in materials science education through real-time feedback.
• To identify critical process stages influencing defect formation.

Main Methods:

• A Long Short-Term Memory (LSTM) network with an Attention mechanism was employed.
• A dataset of 1,000 university kiln firing cycles was analyzed, with 144 time steps per cycle.
• Model performance was evaluated at 25%, 50%, 75%, and 100% completion, comparing against Random Forest and MLP models.

Main Results:

• The LSTM-Attention model achieved 79.3% accuracy at the 50% completion checkpoint, outperforming other models.
• Attention weights highlighted critical temperature zones (quartz transition and sintering phases) associated with defects.
• The model identified problematic intervals, enabling potential for early intervention.

Conclusions:

• Data-driven early warning systems can facilitate real-time guided instruction in ceramic processing.
• Temporal modeling is effective for defect prediction under limited observation.
• Further research is needed to address single-facility data limitations and industrial deployment requirements.