Retrainable physics-integrated neural differentiable modeling of sintering across material systems
Sinter-PiNDiff beat two neural baselines across all twelve held-out material and metric comparisons reported.
The model couples neural networks to sintering rate equations to predict density and grain-size evolution from sparse data. It was separately fitted to published MgO, Al-doped ZnO, and CaO-doped ThO2 datasets using the same structure and training procedure. Removing evolving density from the inputs worsened errors, supporting the paper’s density-feedback claim. Its ensemble uncertainty estimates were useful for disagreement but were not calibrated to cover all data gaps. ArXiv · AI/CL/LG's note
The model couples neural networks to sintering rate equations to predict density and grain-size evolution from sparse data. It was separately fitted to published MgO, Al-doped ZnO, and CaO-doped ThO2 datasets using the same structure and training procedure. Removing evolving density from the inputs worsened errors, supporting the paper’s density-feedback claim. Its ensemble uncertainty estimates were useful for disagreement but were not calibrated to cover all data gaps. ArXiv · AI/CL/LG's note
score 3