Robotic Automation Discovery of Biodegradable Electronics via Multimodal Active Learning and AI-Guided Design

Haochen Yang, Po-Yen Chen

Keywords: Multimodal Dynamic Active Learning, Uncertainty-Aware Sampling, Robotic Autonomous Experimentation, Biodegradable Electronics

TL;DR: This paper presents a multimodal dynamic active learning framework integrated with robotic autonomous experimentation to guide the design and discovery of biodegradable electronics.

Abstract: Designing biodegradable electronic substrates that possess both mechanical robustness and tunable dielectric properties is a complex challenge, characterized by high-dimensional formulation spaces and nonlinear coupling between material properties. To accelerate discovery, we introduce a multimodal dynamic active learning (AL) framework integrated with multiple robotic automation platforms to navigate a vast design space with a 19-component library comprising biopolymers, layered clays, and metal oxides. Unlike static acquisition strategies, this AL framework employs a stage-wise approach that dynamically adapts sampling objectives, shifting from diversity-driven exploration to uncertainty-aware exploitation, to navigate the design space efficiently. This predictive modeling engine is tightly coupled with a fully automated experimental pipeline, featuring robotic liquid handling, automated tensile testing, and dielectric characterization, enabling closed-loop experimentation. Over eight AL rounds, we developed a prediction model leveraging an ensemble of hierarchical multimodal artificial neural networks for simultaneous feasibility screening and property prediction. Benchmarking demonstrates that this dynamic approach yields superior learning efficiency and target discovery compared to baselines under limited experimental budgets. We also successfully identified samples with mechanical robustness and an application-relevant range of dielectric properties demonstrating the practical utility of the closed-loop robotics--ML workflow under realistic experimental constraints.

Submission Track: Feedback-Based Learning for Materials Design - Tiny Paper

Submission Category: AI-Guided Design + Automated Material Characterization

Submission Number: 64