CRESt multimodal AI copilot accelerates materials discovery and delivers record formate fuel cell performance

CRESt: A multimodal copilot for real-world materials discovery

Materials science has a bottleneck: experiment design and iteration take time, money, and endurance. Most machine-learning tools help but operate on narrow data and boxed-in search spaces. The new Copilot for Real-world Experimental Scientists (CRESt) changes that by combining literature insights, compositions, microstructural images, and live experimental feedback with robotics and active learning.

With CRESt, human researchers set goals in natural language, then the system proposes experiments, runs them with lab automation, watches for issues, and updates its hypotheses. It acts like a sharp lab partner that reads, observes, and learns across modalities - then moves samples through synthesis, characterization, and testing without code.

What makes CRESt different

• Multimodal inputs: prior literature, chemical compositions, XRD and SEM images, microstructures, and experimental results.
• Natural-language interface: researchers converse with the system; it explains observations and proposes next steps.
• Active learning + Bayesian optimization (BO) in a reduced space: it builds knowledge embeddings from literature/databases, performs PCA to focus on variables that drive performance, then runs BO in that compact space.
• Self-monitoring experiments: cameras plus visual-language models spot setup drift and procedural errors, flagging corrections in text or voice.
• Robotic throughput: liquid-handling, carbothermal shock synthesis, automated electrochemistry, electron/optical microscopy, and remotely controlled pumps and gas valves.

"In the field of AI for science, the key is designing new experiments," says Ju Li. "We use multimodal feedback - for example information from previous literature on how palladium behaved in fuel cells at this temperature, and human feedback - to complement experimental data and design new experiments."

A faster loop that learns with each run

• Start with prior knowledge: encode literature and databases into recipe embeddings.
• Compress the design space: use principal component analysis to keep what explains most performance variability.
• Choose the next experiment: run Bayesian optimization in the reduced space.
• Automate the lab work: robots synthesize, characterize, and test.
• Watch and correct: computer vision and vision-language models detect anomalies and suggest fixes.
• Close the loop: feed experimental data and human feedback back into a large model; update the reduced space and repeat.

Real results: a record-setting formate fuel cell catalyst

Using CRESt, researchers explored more than 900 chemistries and ran 3,500 electrochemical tests in three months. The system uncovered an eight-element catalyst for a direct formate fuel cell with a 9.3× improvement in power density per dollar over pure palladium.

Despite using just one-fourth the precious metal content of prior devices, the catalyst delivered record power density in a working cell. "We used a multielement catalyst that also incorporates many other cheap elements to create the optimal coordination environment for catalytic activity and resistance to poisoning species," says Zhen Zhang.

Why this matters for your lab

• Search broader, test smarter: move beyond fixed, three-element grids. Let knowledge embeddings and PCA define the space that matters.
• Reduce wasted cycles: use BO to balance exploration and exploitation with every new datapoint.
• Improve reproducibility: camera checks catch subtle deviations (pipette offsets, sample geometry drift) that derail results.
• Keep humans in control: you set objectives, constraints, and priors; the system speeds execution and learning.

Active learning and BO - quick context

Active learning prioritizes which experiment to run next based on what reduces uncertainty or increases expected improvement. Bayesian optimization uses a surrogate model and an acquisition function to propose the next high-value point, which is well-suited to expensive experiments.

• Active learning overview
• Bayesian optimization overview

System components you can adopt now

• Literature-grounded priors: mine papers and databases to build structured priors for recipes and processing conditions.
• Reduced-space modeling: run PCA or a similar technique on knowledge embeddings before BO to avoid getting lost in high dimensions.
• Multimodal data capture: couple SEM/XRD image analysis with electrochemical or mechanical readouts for richer learning signals.
• Experiment monitoring: add low-cost cameras and a vision-language model to detect setup drift and prompt corrections.
• Human-in-the-loop feedback: require short natural-language notes after each run; feed them back into the model.

What stays human

Strategy, constraints, and interpretation. CRESt assists; it doesn't replace domain judgment. The system surfaces options, highlights anomalies, and executes quickly. You decide what to try, what to trust, and where the science goes next.

Upskill your team

If you're building similar AI-assisted workflows for R&D, structured learning can shorten the ramp-up time. See role-based programs here: AI courses by job.

Key takeaways

• Multimodal knowledge plus active learning outperforms single-stream models in complex materials spaces.
• Embedding + PCA before BO focuses search on variables that drive performance.
• Computer vision improves reproducibility by catching small but consequential deviations.
• CRESt delivered an eight-element catalyst with record performance and far lower precious metal usage for direct formate fuel cells.

"CRESt is an assistant, not a replacement, for human researchers," says Ju Li. "This is a step toward more flexible, self-driving labs."