Sample-efficient parameter exploration of the powder film drying process using experiment-based Bayesian optimization

Kohei Nagai, Takayuki Osa, Gen Inoue, Takuya Tsujiguchi, Takuto Araki, Yoshiyuki Kuroda, Morio Tomizawa, Keisuke NagatoScientific 2022 · DOI

TL;DR

Researchers successfully applied Bayesian optimization (BO) to find the "needle in a haystack" optimal drying conditions for powder films, a process critical for manufacturing fuel cells and batteries. By conducting just 40 experiments, the team navigated a massive search space of over 32,000 possible parameter combinations—achieving a 1,000x increase in efficiency over brute-force methods and discovering a counter-intuitive "low-high-low" temperature profile that minimizes film defects.

What Problem Does This Paper Address?

Fabricating high-quality powder films—the thin layers of active material found in fuel cells, solar cells, and batteries—is notoriously difficult. The process involves a complex slurry of powders, polymers, and solvents that must be dried perfectly to avoid cracks and uneven thickness.

Because these phenomena are "multiscale and multiphysical," creating an accurate numerical model to predict the outcome is nearly impossible. Traditionally, engineers rely on "brute force" trial-and-error or "design of experiments" (DoE) methods. However, when you have multiple parameters (like temperature changes over time), the number of possible combinations explodes. In this study, the authors faced 32,768 possible temperature profiles, making traditional exploration prohibitively slow and expensive.

Methods

The authors employed experiment-based Bayesian optimization (BO) to bridge the gap between complex physical phenomena and efficient parameter search.

  1. The Experimental Setup: They used a robot arm and a motorized pipette to drop-cast a slurry (carbon particles and Nafion polymer) onto a heater.
  2. The Parameters: The drying process was divided into five 120-second time steps. For each step, the heater could be set to one of eight temperature levels (30°C to 100°C), resulting in $8^5$ (32,768) possible profiles.
  3. The "Brain" (GPR): They used Gaussian Process Regression (GPR) as a surrogate model. GPR doesn't just predict the "defect ratio" for a given profile; it also estimates the uncertainty of that prediction.
  4. The Strategy (UCB): To decide which experiment to run next, they used an Upper Confidence Bound (UCB) acquisition function. This algorithm balances "exploitation" (testing areas the model thinks are good) with "exploration" (testing areas where the model is uncertain).
  5. Evaluation: They used automated image processing to calculate a "defect ratio"—the percentage of the film covered by cracks or overly thin regions.

Datasets

Unlike many AI papers that use static datasets, the "data" here was generated in real-time through physical experiments:

  • Initial Dataset: 30 experiments conducted with randomly selected temperature profiles to give the BO system a baseline.
  • Optimization Dataset: 10 additional experiments directed by the BO algorithm.
  • Total: 40 experimental data points used to navigate a search space of 32,768 candidates.

Key Insights and Ideas

  • The "Low-High-Low" Discovery: The most surprising finding was that the optimal drying profile wasn't a simple constant temperature or a steady ramp. Instead, the system favored a profile that started low (30°C), jumped high (>60°C), and then dropped back down.
  • The "Skin" Mechanism: The authors inferred a physical mechanism for this success: the initial low temperature prevents the "coffee-ring effect" (uneven thickness). The subsequent high-temperature spike forms a "skin" of carbon particles on the surface, which then regulates the remaining evaporation to prevent the internal stresses that cause cracks.
  • Efficiency over Intuition: This specific "process window" is extremely narrow. A human researcher using traditional methods would likely have missed it, as it contradicts the standard industrial practice of using a single, steady drying temperature.

Key Results

  • Massive Speedup: The optimization converged in just 40 experiments, representing only 0.1% of the total parameter space. This is roughly 1,000 times more efficient than a brute-force search.
  • Defect Reduction: While random profiles often resulted in defect ratios between 30% and 50%, the BO-guided search quickly found profiles with a defect ratio as low as 2.8%.
  • Convergence: The system showed clear convergence; by the 10th BO-guided experiment, the predicted defect ratios were consistently low and the model's uncertainty had significantly decreased.

Limitations and Future Work

  • Generalizability: The study focused on a specific slurry (carbon/Nafion) and a specific method (drop casting). It remains to be seen how well these specific "low-high-low" rules apply to other materials or industrial-scale coating methods like slot-die coating.
  • Evaluation Subjectivity: While the authors used image processing to be quantitative, the "defect ratio" is a proxy for performance. Future work could integrate direct electrochemical performance measurements (like power density in a fuel cell) into the BO loop.
  • Process Informatics: The authors propose this as a step toward "process informatics"—a field that applies the data-driven lessons of materials informatics to the actual manufacturing steps, which are often the true bottleneck in bringing new materials to market.

Summary updated

Abstract

Parameter optimization is a long-standing challenge in various production processes. Particularly, powder film forming processes entail multiscale and multiphysical phenomena, each of which is usually controlled by a combination of several parameters. Therefore, it is difficult to optimize the parameters either by numerical-model-based analysis or by "brute force" experiment-based exploration. In this study, we focus on a Bayesian optimization method that has led to breakthroughs in materials informatics. Specifically, we apply this method to exploration of production-process-parameter for the powder film forming process. To this end, a slurry containing a powder, polymer, and solvent was dropped, the drying temperature and time were controlled as parameters to be explored, and the uniformity of the fabricated film was evaluated. Using this experiment-based Bayesian optimization system, we searched for the optimal parameters among 32,768 (8<sup>5</sup>) parameter sets to minimize defects. This optimization converged at 40 experiments, which is a substantially smaller number than that observed in brute-force exploration and traditional design-of-experiments methods. Furthermore, we inferred the mechanism corresponding to the unknown drying conditions discovered in the parameter exploration that resulted in uniform film formation. This demonstrates that a data-driven approach leads to high-throughput exploration and the discovery of novel parameters, which inspire further research.
BibTeX
@inproceedings{nagai2022sample,
  author = {Kohei Nagai and Takayuki Osa and Gen Inoue and Takuya Tsujiguchi and Takuto Araki and Yoshiyuki Kuroda and Morio Tomizawa and Keisuke Nagato},
  title = {{Sample-efficient parameter exploration of the powder film drying process using experiment-based Bayesian optimization}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-022-05784-w},
}
Figure