From Limited Data to Scalable Discovery: A Robotics and Machine Learning Integrated Workflow for Accelerated Materials Innovation

The development of next-generation functional materials is often constrained by small, high-variance datasets and the inefficiency of labor-intensive, trial-and-error experimentation. To overcome these barriers, we present a robotics–machine learning integrated workflow that transforms limited data into scalable discovery. Automated robotic platforms enable high-throughput and reproducible experimentation, rapidly generating diverse datasets across high-dimensional design spaces. Machine learning strategies, incorporating active learning and data augmentation, construct ensemble-based prediction models that deliver accurate and robust performance even under data scarcity. By coupling these models with interpretability tools, the workflow expands accessible property boundaries and accelerates the optimization of functional materials with tunable properties. As a case study, we demonstrate this approach in the discovery of sustainable biobased nanocomposite packaging films. An automated pipetting robot formulated 2,420 nanocomposites, whose film quality data trained an artificial neural network classifier to define a design space. Within this space, 16 active learning loops iteratively fabricated and characterized 343 nanocomposites, producing a high-quality dataset. Leveraging this dataset alongside density functional theory simulations, a prediction model explored ~1 billion formulations, identifying candidates with superior mechanical resilience and tunable transparency. Among them, Cu2+-incorporated nanocomposite films exhibited moisture absorption, oxygen impermeability, and antimicrobial performance, outperforming conventional plastics and extending the shelf life of postharvest produce. Life cycle assessment–informed feedback further refined the formulations, and a data-sharing platform was established to promote adoption. This case study highlights the power of the robotics–machine learning workflow in accelerating materials discovery from limited, high-variance datasets, offering a scalable pathway for the predictive design of advanced and sustainable materials. Beyond biobased packaging, we have extended this closed-loop robotics and machine learning workflow to a range of materials challenges, including soft electronics, conductive aerogels, smart soft robotics, stretchable conductors, and battery electrolytes. Across these domains, the core idea is the same: reproducible automated experimentation paired with data-efficient, interpretable models to rapidly map the design space and converge on high-performing formulations under data scarcity.