Augmented Experiment in Material Engineering Using Machine Learning

Aomar Osmani, Massinissa Hamidi, Salah BouhoucheAAAI 2021 · DOI

Introduction

The discovery and synthesis of new materials, such as anticorrosive paints derived from industrial by-products like calamine, is a cornerstone of modern material engineering. However, the traditional experimental pipeline is often hindered by high costs and significant time requirements. A primary tool in this process is Thermogravimetric Analysis (TGA), which tracks changes in a material's mass and heat flow as a function of temperature. While TGA provides critical insights into phase transitions and chemical stability, performing exhaustive physical experiments for every possible mixture ratio is economically unfeasible.

Figure 1: Typical DSC-TGA analysis of a red pigment sample, showing weight loss and heat flow curves.

The Modeling Challenge

Material scientists face a dilemma when modeling these complex systems. Purely analytical models, based on established physical equations, often fail to capture the non-linear "black box" dynamics of novel mixtures. Conversely, standard Machine Learning (ML) models, while flexible, frequently ignore fundamental physical laws like mass conservation. This lack of "physical grounding" leads to poor generalization, especially when data is sparse or when the model is required to extrapolate beyond the training set.

Proposed Methodology: Physics-Informed Machine Learning

To bridge this gap, the authors propose a hybrid framework that combines the predictive power of neural networks with the theoretical rigor of classical kinetic models.

Hybrid Architecture and Regularization

The core of the approach is a neural network with three fully connected layers using ReLU activation. To ensure physical consistency, the researchers incorporate Arrhenius and Eyring kinetic equations directly into the loss function. These equations describe the degradation kinetics of chemical reactants, effectively "circumscribing" the neural network's behavior within theoretically sound boundaries.

Optimization via Conditional Gradient

Rather than relying solely on standard Stochastic Gradient Descent (SGD), which may prioritize data-fitting at the expense of physical constraints, the study employs Conditional Gradient (CG) optimization. This method is used to find Pareto-optimal solutions that satisfy both empirical data and physical laws simultaneously, ensuring the model remains physically plausible even in regions where experimental data is missing.

Experimental Setup and Datasets

The model was validated using an industrial dataset of red pigment (iron ore) and calamine (a steel industry by-product) mixtures. Experiments were conducted using an SDT-Q600 instrument, capturing 3,000 measurement points per run across various calamine percentages (5% to 35%). The goal was to evaluate the model's ability to reconstruct the material's thermal signature both through interpolation (inside known regions) and extrapolation (outside known regions).

Key Results and Discussion

The hybrid approach demonstrated significant improvements over baseline ML models. Notably, the model achieved a substantial reduction in perplexity for weight reconstruction (2.76 vs. 3.29).

Figure 6: Mean Squared Error (MSE) trends for interpolation (inside) and extrapolation (outside) across different validation distances.

As shown in Figure 6, the model maintains controlled error rates even when extrapolating up to 30% away from the training data points. This robustness is a direct result of the "inductive bias" introduced by the kinetic-based regularization. By enforcing mass conservation and kinetic bounds, the framework prevents the model from generating "unphysical" predictions that a standard neural network might produce in data-poor regimes.

Conclusion

This research presents a viable path toward "augmented experiments" in material science. By integrating domain-specific physical knowledge with machine learning, engineers can significantly reduce the number of costly TGA runs needed to characterize new materials. While limitations remain—such as the loosening of penalty bounds during extreme extrapolation—the hybrid framework offers a powerful tool for accelerating material discovery and optimizing industrial processes.

Summary updated

BibTeX
@inproceedings{osmani2021augmented,
  author = {Aomar Osmani and Massinissa Hamidi and Salah Bouhouche},
  title = {{Augmented Experiment in Material Engineering Using Machine Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i10.17116},
}
Figure