AI for Experiment Planning and Control

Ducci and Kunkel Group

Welcome to the AI for Experiment Planning and Control group of the FHI Theory Department. We work at the intersection of agentic AI, machine learning, experimental automation, and physical modeling for accelerating scientific discovery in catalysis and solar battery materials.

Our Work

Instead of merely screening for “one more optimal material,” our core focus is the extraction of fundamental mechanistic insights in the catalysis field. We develop robust, chemically interpretable machine learning models, including sparse‑regression, symbolic‑regression, and physics‑informed frameworks. By embedding physical principles directly into the mathematical architecture, we transform traditional black‑box surrogates into transparent, mechanistically suggestive models. This strategy allows us to decode complex, non-linear chemical systems where simulation-only approaches currently fall short.

These models are not only tools for interpretation; they also provide the decision-making foundation for automated, closed-loop experimental workflows. Moving beyond passive data analysis, we propose data-efficient adaptive design of experiments (DoE) frameworks to actively steer laboratory campaigns, autonomously determining the next optimal measurement based on real-time feedback. This allows us to efficiently navigate high‑dimensional spaces, driving targeted discovery with minimal experimental overhead.

From algorithms to the lab

The developed workflows extend beyond the conceptual stage and are already running on physical laboratory infrastructure.

• Automated Electron Microscopy – Smart control of environmental scanning electron microscopy (ESEM) experiments, paired with automated image segmentation and quantitative analysis of catalyst surface dynamics. 

• Active Reactor Profiling – Real-time selection and optimization of operating conditions for catalytic reactors, guided by on-the-fly model updates and adaptive experimental planning.

• Novel catalyst discovery – Systematic exploration of multi‑promoter composition spaces using data‑efficient, interpretable adaptive DoE, revealing interactions, trade‑offs, and Pareto‑optimal formulations that outperform industrial benchmarks.

We build our closed-loop applications on top of a dedicated infrastructure stack. Our setups utilize a modular, EPICS-based control architecture that integrates physical hardware—including microscopes, reactors, and robotic handlers—into a single framework. This architecture natively incorporates FAIR data management and is designed to support a federated infrastructure across different laboratories.

We deploy this framework directly within our collaborations. We contribute the core theory, AI, and software/hardware tools to the MPG–TUM Center for Solar Batteries and Optoionic Technologies (SolBat) for the study of optoionic materials. Within the SusMax project we use our data-driven tools to investigate the catalytic conversion of renewable feedstock. Furthermore, we are a key partner in ASCEND, a BMFTR-funded initiative building the next generation of self-driving laboratories for catalyst discovery, where we help develop the central decision-making “brain.”

Vision
The long‑term vision is a laboratory that autonomously designs experiments, interprets the resulting measurements, revises its mechanistic hypotheses, and decides what to test next — with adaptive, interpretable planners acting as the core decision‑making module.

Selected Publications

Selected recent publications are listed below (a full list can be found under Publications):

Physics-Based Synthetic Data Model for Automated Segmentation in Catalysis Microscopy, M. Vuijk et al.,  Microsc. Microanal. (2024)

Model driven adaptive design with concentration profiles, M. Kouyaté et al., J. Chem. Phys. (2025)

Pareto-based optimization of sparse dynamical systems, G. Ducci, M. Kouyaté, K. Reuter, and C. Scheurer, J. Chem. Phys. (2025)

Systematic Exploration of a Multi-Promoter Catalyst Composition Space with Limited Experiments: Non-Oxidative Propane Dehydrogenation to Propylene, C. Kunkel et al., ACS Catal. (2024)

Adaptive Experiment Planning for Inverse Design and Understanding: Synergistic Interactions as Key to Optimized Multi-Promoter Formulations, C. W. P. Pare et al., ACS Catal. (2026)