🔬 materials science

Retrainable physics-integrated neural differentiable modeling of sintering across material systems

The paper introduces Sinter-PiNDiff, a retrainable physics-integrated neural differentiable framework that accurately predicts sintering density and grain-size evolution across diverse material systems with sparse data, outperforming standard neural network baselines while highlighting the critical role of density-dependent kinetic feedback.

Original authors: Zeping Chen, Ani Aprahamian, Khachatur V. Manukyan, Tengfei Luo

Published 2026-09-28

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Zeping Chen, Ani Aprahamian, Khachatur V. Manukyan, Tengfei Luo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Ceramics are the silent workhorses of modern technology, found in everything from the insulators on power lines to the components inside jet engines and the chips in our computers. To make these materials useful, manufacturers must turn loose powders into solid, dense blocks through a process called sintering. Imagine taking a pile of fine sand and heating it until the individual grains fuse together, shrinking the gaps between them to form a single, strong piece. This heating process is a delicate balancing act. As the material heats up, the tiny pores between the grains disappear, making the object denser and stronger. At the same time, the grains themselves grow larger, merging with their neighbors. If the grains grow too much, they can trap the remaining pores inside, leaving the material weak and porous. If they don't grow enough, the material might not reach its full strength. For decades, engineers have had to rely on trial and error, running countless furnace tests to find the perfect temperature and time for each specific type of ceramic, because predicting exactly how a material will behave is incredibly difficult.

A team of researchers has developed a new way to solve this puzzle, creating a digital tool that can predict how different ceramics will behave during sintering without needing to run a single physical experiment first. The tool, which they call Sinter-PiNDiff, acts like a smart guide that learns from limited data to understand the complex dance of shrinking and growing that happens inside the kiln. Unlike older computer models that require a massive amount of specific data for every single material, or simple machine learning programs that try to guess the outcome without understanding the underlying physics, this new approach blends the two. It uses the known laws of how materials shrink and grow, but it lets a computer program learn the specific details for each new material it encounters. The researchers tested this system on three very different types of ceramics: magnesium oxide, aluminum-doped zinc oxide, and calcium-doped thorium oxide. These materials have different chemical makeups and behave in distinct ways, yet the same digital framework was able to learn and predict their behavior accurately in each case.

The core of this achievement lies in how the system handles the relationship between density and grain size. In the real world, as a ceramic part gets denser, the way its grains grow changes, and as the grains grow, they affect how fast the part can get denser. The new model captures this two-way conversation. It uses a set of equations that describe the general rules of sintering, but it fills in the missing pieces with a neural network—a type of artificial intelligence that can learn patterns from data. As the computer simulates the heating process step by step, it constantly updates its predictions based on the current state of the material. If the material is getting denser, the model adjusts its expectations for how the grains will grow, and vice versa. This allows the system to learn the specific "personality" of a material, such as how quickly it densifies or how much its grains grow at a certain temperature, using only a few scattered measurements from past experiments.

When the researchers put their model to the test, it outperformed other common computer methods. They compared their results against two other types of artificial intelligence: one that simply tries to draw a line through the data points without understanding the process, and another that tries to learn the rate of change at every step. The new model was more accurate in predicting both the final density and the final grain size for all three materials. For magnesium oxide, the model predicted the density with an error of about 14.6 percent, and the grain size with an error of about 8.6 percent. For the aluminum-doped zinc oxide, the density error was even lower at 10.8 percent, and for the calcium-doped thorium oxide, it was 14.4 percent. These numbers are significant because they show that the model can generalize well, making accurate predictions even for temperatures and chemical compositions it had never seen before. For instance, when asked to predict the behavior of a material with a high concentration of calcium that was not part of the training data, the model still provided a useful estimate, though it correctly signaled that it was less certain about this specific case.

The study also revealed that feeding the model information about the current density of the material was crucial for its success. When the researchers removed this piece of information, forcing the model to guess the density's effect without actually knowing it, the predictions became less accurate. This confirms that the changing state of the material is a key driver of the process and that the model needs to "see" the material as it evolves to make good predictions. The system also includes a way to show how confident it is in its answers. By running the simulation multiple times with slightly different starting points, the model produces a range of possible outcomes. When the data is sparse or the conditions are very different from what it has seen, this range widens, alerting the user that the prediction is less certain. This feature is vital for engineers, as it tells them exactly where they might need to run a new physical experiment to get better data.

Ultimately, this work offers a practical path forward for designing better ceramics. Instead of relying on expensive and time-consuming furnace trials to figure out the right settings for a new material, engineers can use this tool to narrow down the possibilities quickly. The model can suggest which temperatures and times are likely to produce the strongest, densest parts, and it can highlight which conditions are too risky to guess at without more data. While the model is not a crystal ball that can predict every microscopic detail, it provides a reliable, physics-based map for navigating the complex world of ceramic manufacturing. By combining the reliability of physical laws with the flexibility of machine learning, the researchers have created a tool that can be retrained for new materials, opening the door to faster development of the advanced ceramics needed for energy, electronics, and structural applications.

Technical Summary: Retrainable Physics-Integrated Neural Differentiable Modeling of Sintering

Problem Statement
Sintering is a critical process for manufacturing ceramics, yet predictive modeling remains challenging due to the coupled nature of densification and grain growth, material-dependent kinetics, and the scarcity of experimental data. Traditional process design relies on costly, iterative furnace trials because existing physics-based models require fixed parameters that often fail to generalize across different ceramic chemistries and microstructures. Conversely, purely data-driven machine learning models struggle to learn temporal evolution from sparse datasets where complete time histories are unavailable for many temperatures or compositions. There is a need for a framework that can integrate physical laws with data-driven learning to predict density and grain-size evolution across diverse material systems using limited measurements.

Methodology: Sinter-PiNDiff
The authors propose Sinter-PiNDiff, a retrainable physics-integrated neural differentiable framework. The core innovation lies in embedding neural networks within a set of coupled sintering rate equations rather than replacing the physics entirely.

Governing Equations: The model retains a physical backbone based on classical sintering relations where the rate of relative density change ( $\dot{\rho}$ ) and grain growth ( $\dot{G}$ ) are functions of temperature, current density, grain size, and composition. The equations are:
$\dot{\rho} = C_0 \psi(\rho) \frac{D_{nn}}{G^m RT}$
$\dot{G} = \frac{k_{nn}}{G^p}$
Here, $m=3$ and $p=2$ are fixed exponents based on volume diffusion and bulk transport scaling. A smooth density-saturation factor, $\psi(\rho) = \tanh((1-\rho)/0.05)$ , attenuates densification as theoretical density is approached.
Neural Kinetic Closures: The unresolved kinetic factors (activation energies, prefactors, and microstructural contributions) are grouped into two effective coefficients: $D_{nn}$ (effective densification) and $k_{nn}$ (effective grain growth). Two fully connected neural networks learn these coefficients. Crucially, the networks take the current density ( $\rho_n$ ), temperature ( $T$ ), and composition ( $\chi_d, \chi_b$ ) as inputs, allowing the kinetic coefficients to evolve dynamically as the material state changes.
Differentiable Autoregressive Training: The system is implemented as a single differentiable program using explicit Euler integration. This allows gradients to propagate through the entire time rollout, enabling the model to be trained end-to-end using sparse observations of density and grain size at various time points.
Uncertainty Quantification: To address epistemic uncertainty arising from sparse data, the authors employ Deep Ensembles (five independently initialized models). The spread of predictions among these members serves as a measure of model disagreement.
Evaluation Strategy: The framework was tested on three distinct ceramic systems: MgO (ionic, single composition), Al-doped ZnO (mixed bonding, multiple dopant levels), and CaO-doped ThO2 (ionic with vacancy modification). The study utilized a "condition-wise" train-test split, withholding entire temperature-composition trajectories to test interpolation and extrapolation capabilities.

Key Results

Predictive Accuracy: Sinter-PiNDiff achieved the lowest mean error in all twelve material–metric comparisons against two data-driven baselines: a Multilayer Perceptron (MLP) and a Residual Network (ResNet).
- Density NRMSE: 14.6% (MgO), 10.8% (Al-doped ZnO), and 14.4% (CaO-doped ThO2).
- Grain-Size NRMSE: 8.6% (MgO), 12.1% (Al-doped ZnO), and 19.3% (CaO-doped ThO2).
- The physics-integrated approach produced smoother, more physically consistent temporal trajectories compared to the baselines, which often exhibited non-monotonic behavior or excessive ensemble spread.
Generalization: The model successfully generalized to unseen temperatures and compositions. It demonstrated capability in interpolating between training conditions and, to a lesser extent, extrapolating to higher dopant concentrations (e.g., 8.8 mol% CaO in ThO2), though extrapolation showed increased uncertainty and systematic bias.
Ablation Study: Removing the evolving density ( $\rho_n$ ) from the neural network inputs increased trajectory errors in all three systems for ten of the twelve metrics examined. This confirms that the learned kinetic coefficients depend on the evolving material state (density-dependent kinetic feedback).
Uncertainty Analysis: While deep ensembles provided a measure of disagreement, the empirical coverage of the $\pm 3\sigma$ bands was not calibrated (ranging from 47% to 90% depending on the system and observable). The authors note that ensemble spread reflects disagreement among models trained on finite data rather than a guaranteed confidence interval.

Significance and Claims
The paper claims that Sinter-PiNDiff establishes a retrainable framework capable of predicting coupled sintering behavior across chemically distinct oxide systems using sparse literature data. By separating a common physical structure from material-specific kinetics, the framework avoids the need for separate empirical formulations for each system.

The authors position the work as a step toward uncertainty-informed experimental design. While the model cannot identify specific microscopic mechanisms (e.g., specific diffusion pathways or secondary phase distributions) from density and grain size alone, the composition-dependent changes in predicted trajectories can highlight conditions where additional characterization (e.g., X-ray diffraction or microscopy) would be most informative. The framework is presented as a tool for screening processing conditions and guiding the selection of sintering schedules where data is scarce, rather than as a definitive solver for all sintering regimes. The authors remain modest regarding extrapolation, noting that performance degrades when testing conditions involve kinetic regimes insufficiently represented in the training data.

Technical Summary: Retrainable Physics-Integrated Neural Differentiable Modeling of Sintering

More like this