Assistant Professor Jie Xu of the University of Chicago Pritzker School of Molecular Engineering, who also holds a joint appointment at Argonne National Laboratory, is part of a growing group of researchers rethinking how artificial intelligence should be used in experimental science. Rather than pushing laboratories toward full automation, Xu and colleagues are developing systems that deliberately keep humans involved in key decisions, even as AI takes on a central analytical role.
Dai, Y., Chan, H., Vriza, A., Fan, J., Kim, F., Wang, Y., Liu, W., Shan, N., Xu, J., Weires, M., Wu, Y., Cao, Z., Miller, C. S., Divan, R., Gu, X., Zhu, C., Wang, S., & Xu, J. (2025). Adaptive AI decision interface for autonomous electronic material discovery. Nature Chemical Engineering, 2(12), 760–770. https://doi.org/10.1038/s44286-025-00318-3
So-called self-driving or autonomous laboratories use machine learning to design experiments, analyze results, and decide what to test next. These platforms are becoming more common in materials science, where the number of possible material combinations quickly exceeds what humans can explore manually. At the same time, they have raised questions about whether machines should be allowed to operate independently in scientific discovery. The work from Argonne and the University of Chicago proposes a middle path.
Assistant Professor Jie Xu of the University of Chicago Pritzker School of Molecular Engineering stated,
“While AI is excellent at this form of data analysis, it falters at decision-making when there are few data points to guide it”.
In a recent study published in Nature Chemical Engineering, the team introduced what they describe as an AI advisor. Instead of acting as the sole decision-maker, the system continuously monitors experimental progress, evaluates performance, and provides recommendations to human researchers. The final decisions remain with the scientists, who can choose to adjust strategies, refine the design space, or continue along the current path.
The approach is inspired by advisory software used in financial trading, where algorithms process large volumes of data in real time but do not replace human judgment. In the lab setting, the AI advisor analyzes incoming experimental data and flags situations where progress slows or results diverge from expectations. This allows researchers to intervene early, rather than committing to a single automated strategy from start to finish.
Henry Chan, a staff scientist in Argonne’s nanoscience and technology division and a co-corresponding author on the study, describes the goal as collaboration rather than competition between humans and machines. AI excels at pattern recognition and rapid analysis, while experienced researchers are better at reasoning under uncertainty, especially when data are sparse. The advisor model is designed to let each side focus on its strengths.
To test the concept, the researchers deployed the system in Polybot, a self-driving lab located at Argonne’s Center for Nanoscale Materials. The platform was used to study and optimize a mixed ion-electron conducting polymer, a class of materials relevant to energy storage and electronic devices. These materials are challenging to design because their performance depends on a combination of structural and processing parameters that interact in complex ways.
Using the AI advisor framework, the team reported a substantial improvement in material performance compared with previous automated approaches. The optimized polymer showed significantly higher mixed conducting capability, and the experiments helped identify two structural features that played a key role: increased spacing between crystalline layers and a larger specific surface area. These insights provide guidance for future materials design beyond the specific system studied.
Sihong Wang, an associate professor at UChicago PME and another co-corresponding author, emphasizes that materials research often has two parallel goals. One is to improve performance, and the other is to understand why that improvement occurs. By allowing researchers to explore a wider range of design variations while staying engaged in decision-making, the AI advisor helped advance both objectives at the same time.
The study also highlights a limitation of current AI systems. While they perform well when large datasets are available, they struggle to make reliable decisions early in an experiment, when only a few results exist. Human intuition and experience remain critical at this stage. The advisor model is intended to bridge this gap rather than eliminate it.
Looking ahead, the team sees opportunities to deepen the interaction between humans and AI. At present, most of the information flows from the AI advisor to the researcher, who then decides how to respond. Future versions could allow the system to learn from human interventions, gradually incorporating elements of human decision-making into its own models.
The broader significance of the work lies in its generality. The researchers argue that the advisor framework could be adapted to many types of self-driving labs, not only in electronic materials but across chemistry, materials science, and engineering. Rather than framing autonomy as a question of who is in control, the study suggests that shared control may be a more productive way to accelerate discovery while preserving scientific insight.
Adrian graduated with a Masters Degree (1st Class Honours) in Chemical Engineering from Chester University along with Harris. His master’s research aimed to develop a standardadised clean water oxygenation transfer procedure to test bubble diffusers that are currently used in the wastewater industry commercial market. He has also undergone placments in both US and China primarely focused within the R&D department and is an associate member of the Institute of Chemical Engineers (IChemE).
