"AI advisor" helps scientists steer autonomous labs
Autonomous labs are moving from concept to daily practice. A team from Argonne National Laboratory and the University of Chicago Pritzker School of Molecular Engineering (UChicago PME) proposes a simple idea with big impact: let AI and humans share control via an "AI advisor."
Inspired by investment software, the advisor monitors progress in real time, analyzes data, and flags when performance drops. Researchers keep the decision-making role-choosing whether to switch strategy, refine the design space, or update constraints based on the advisor's prompts.
How the advisor works
UChicago PME Asst. Prof. Jie Xu, who holds a joint appointment at Argonne, describes the advisor as a bridge between automation and expert judgment. It brings fast, consistent analysis to every experimental step while keeping scientists in the loop when tradeoffs matter.
Argonne staff scientist Henry Chan adds that the point isn't to crown a winner between humans and AI. It's to divide the work cleanly: AI handles data processing and progress tracking; researchers apply intuition, context, and real-time decisions. "We promote human-machine collaboration to boost discovery together."
Field test: Polybot and mixed ion-electron conducting polymers
The team deployed the advisor on Polybot, a self-driving lab at Argonne's Center for Nanoscale Materials, to design a mixed ion-electron conducting polymer. The target: better transport performance and clearer design rules.
The result was a 150% improvement in mixed conduction versus a previous state-of-the-art approach. Just as important, the workflow surfaced two key factors linked to that improvement-evidence that the method speeds performance gains while sharpening mechanistic insight.
As UChicago PME Assoc. Prof. Sihong Wang notes, materials research has twin goals: raise performance and explain why it improves. By widening the space of structural variations and guiding testing efficiently, the advisor helped achieve both in one pass.
Why this matters for your lab
- Set clear KPIs: define the metrics the advisor should track (yield, stability, transport metrics, sample efficiency).
- Use triggers, not overrides: let the advisor prompt human review when performance stalls or drifts, rather than auto-pivoting.
- Keep a decision log: record advisor prompts and human actions to accelerate learning across iterations and teams.
- Inject domain priors: constrain design spaces with known chemistries, safety rules, and manufacturability.
- Validate across conditions: confirm gains hold across batches, instruments, and environmental ranges.
- Plan for two-way learning: allow the AI to learn from expert actions so it models real decision strategies over time.
What's next
The team's next step is tighter integration-where the AI learns from scientists' choices and updates how it prioritizes experiments. That loop should further reduce trial-and-error, especially in large design spaces.
Details are in "Adaptive AI decision interface for autonomous electronic material discovery" (Dai et al.), published in Nature Chemical Engineering on December 18, 2025. Read the paper via the DOI: 10.1038/s44286-025-00318-3.
For teams building similar capabilities
If you're upskilling your group on AI decision interfaces, experimentation workflows, or lab automation, see curated training paths by role: Complete AI Training - Courses by Job.
Your membership also unlocks:
