Autonomous Experiments in Practice: EPICS, Ophyd, and AI Convergence

Modern experimental workflows in the laboratory often involve precise, synchronized control and monitoring of a diverse set of instruments, including pressure sensors, power supplies, motor controllers, and imaging systems. Even a relatively straightforward task—such as heating a sample inside an ultra-high vacuum (UHV) chamber—requires tight coordination of multiple devices: controlling current through the sample while simultaneously monitoring chamber pressure, temperature, and vacuum conditions in real time. To enable such high-precision experimental protocols, we utilize EPICS-Ophyd-Bluesky framework. EPICS provides a robust, scalable backend for real-time device communication and control and is the first step towards automation of Lab. Various clients, including LabVIEW, Phoebus, and Python-based interfaces, can be used to interact with EPICS IOCs. On top of this, the Ophyd layer offers a Pythonic abstraction for representing physical devices as software objects, allowing for efficient parallel coordination and control of complex device networks. Combined with the Bluesky data acquisition framework, this stack enables the design, execution, and logging of complex, reproducible experiments in a modular and scriptable environment. The digitalized and modular control architecture opens the door for meaningful AI integration—both in real-time experiment orchestration and offline data analysis. AI models can leverage this structured framework to enhance experimental decision-making in the following ways:

· Closed-loop Experiment Control: AI agents can monitor live sensor data and autonomously adjust control parameters (e.g., heating rate, pressure thresholds) to maintain optimal experimental conditions or react to unexpected dynamics in real time.

· Intelligent Experiment Planning: Based on prior datasets, AI models (e.g., Bayesian optimization, reinforcement learning) can propose new experiment configurations that are more likely to yield novel or desirable results, significantly accelerating discovery.

· Automated Data Analysis and Feedback: Machine learning techniques can be applied to experimental outputs—such as spectroscopy, imaging, or time series data—to extract key features, classify patterns, or detect anomalies, and feed this insight back into the experimental pipeline.