Active Learning and Explainable AI for Multi-Objective Optimization of Spin Coated Polymers

TL;DR

The paper tackles the challenge of optimizing spin-coated polymers with competing mechanical properties by integrating active Pareto front learning (PyePAL) with explainable AI tools. Gaussian processes map processing variables $S$, $d$, and concentrations $c_{ ext{PVP10}}$, $c_{ ext{PVP40}}$, $c_{ ext{PVP360}}$ to hardness and elasticity and drive adaptive sampling toward the $\epsilon$-Pareto front, while UMAP provides low-dimensional visualizations and fuzzy linguistic summaries translate learned relationships into interpretable statements. Experimental results demonstrate efficient Pareto front identification and interpretable explanations, enabling expert-driven knowledge discovery in materials optimization. The approach advances transparent AI-assisted materials optimization and can be extended to other high-dimensional processing spaces and properties.

Abstract

Spin coating polymer thin films to achieve specific mechanical properties is inherently a multi-objective optimization problem. We present a framework that integrates an active Pareto front learning algorithm (PyePAL) with visualization and explainable AI techniques to optimize processing parameters. PyePAL uses Gaussian process models to predict objective values (hardness and elasticity) from the design variables (spin speed, dilution, and polymer mixture), guiding the adaptive selection of samples toward promising regions of the design space. To enable interpretable insights into the high-dimensional design space, we utilize UMAP (Uniform Manifold Approximation and Projection) for two-dimensional visualization of the Pareto front exploration. Additionally, we incorporate fuzzy linguistic summaries, which translate the learned relationships between process parameters and performance objectives into linguistic statements, thus enhancing the explainability and understanding of the optimization results. Experimental results demonstrate that our method efficiently identifies promising polymer designs, while the visual and linguistic explanations facilitate expert-driven analysis and knowledge discovery.

Figures (7)

• Figure 1: On the left is the $\epsilon$-PAL workflow, where a Gaussian process model is iteratively updated to map the design space (e.g., spin coating parameters such as spin speed, dilution, and polymer concentration) to the objective space (e.g., nanoindentation measurements such as elasticity and hardness). Points are sampled to minimize the model's uncertainty. On the right, a FLS of the optimization data is constructed, evaluated, and simplified based on user preferences.
• Figure 2: Iteration 7 of PyePAL on the Binh-Korn benchmark. (a) Design Space; (b) Criteria Space. The few black points are those identified as Pareto optimal, orange points are those currently discarded, and green points are unknown.
• Figure 3: Iteration 14 of PyePAL on the Binh-Korn benchmark. (a) Design Space; (b) Criteria Space. The complete Pareto front is identified in black, orange points are those currently discarded, and there are no undecided points.
• Figure 4: Human expert-developed projection of the spin coating output space of (inverse) elasticity. 5 input parameters are compressed in a 2D figure. Making such visualizations costs a lot of time and will ultimately fail when more experimental parameters are added. A more flexible approach of projecting the space is necessary.
• Figure 5: Optimization data from PyePAL after the 5th iteration. The pareto front has not fully been identified, but FLS statements can already be made about the model that guides the algorithm.