Retrainable physics-integrated neural differentiable modeling of sintering across material systems
Sinter-PiNDiff predicts ceramic density and grain growth from sparse data across three material systems.
Sinter-PiNDiff uses two neural networks inside coupled rate equations to predict density and grain-size evolution during ceramic sintering, with a saturation factor near theoretical density. The same architecture was fitted independently to published MgO, Al-doped ZnO, and CaO-doped ThO2 data and beat multilayer perceptron and residual-network baselines on all twelve held-out comparisons. Density normalized RMSE was 14.6%, 10.8%, and 14.4%, and grain-size errors were 8.6%, 12.1%, and 19.3%. Deep ensembles quantified disagreement, but the uncertainty bands were not calibrated.
- Networks learn densification and grain-growth coefficients inside coupled rate equations.
- Fitted separately to published MgO, Al-doped ZnO, and CaO-doped ThO2 data.
- Lowest error versus MLP and ResNet on all twelve comparisons.
- Density NRMSE was 14.6%, 10.8%, and 14.4% for the three systems.
- Ensemble uncertainty bands were uncalibrated and missed some discrepancies.
- Tags
- physics-informed-neural-networkssinteringdifferentiable-modelingceramicsmaterials-scienceuncertainty
- AI models
- Sinter-PiNDiff
From the original article
1 min read at arxiv.orgSintering is widely used to manufacture ceramics, but coupled densification and grain growth, material-dependent kinetics, and sparse measurements complicate predictive modeling and process design. We present Sinter-PiNDiff, a retrainable physics-integrated neural differentiable framework for predicting density and grain-size evolution.
By Zeping Chen · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.