A new project led by the Department of Energy's SLAC National Accelerator Laboratory brings together scientists and engineers from national labs, universities and industry to build a platform that transforms today's complex research environment into self-driving laboratories controlled by Super Intelligence (SI) systems that can evaluate and respond autonomously to rapidly changing needs.
SPIRE: A Source-to-Discovery Platform for Instrumentation, Robotics and Embodied AI in DOE Photon Science Facilities, was announced today by the DOE Office of Science as part of a nationwide effort to create robotics and automation testbeds for scientific discovery. The announcement was made during the Golden Age of Science Summit in Washington, D.C., where a new batch of Genesis Mission projects were also unveiled. SLAC will partner on SPIRE with Stanford University, the University of Chicago, Argonne National Laboratory, Brookhaven National Laboratory, Lawrence Berkeley National Laboratory, and advisors from industry.
Experiments at SLAC and other DOE facilities examine nearly every kind of sample imaginable - from fusion fuel targets to water molecules, plant stems to proteins, sturdy battery cells to ancient manuscripts. SPIRE will help accelerate discoveries at the pace needed to meet the complex and urgent needs of our time.
The self-driving laboratory testbed will touch every part of SLAC's large-scale X-ray and electron user facilities to sharpen the precision of experiments, automatically adjust X-ray and electron beams, and slash the timespan between collecting scientific data and understanding it.
The project aligns with the U.S. effort to create autonomous laboratories outlined in the DOE Genesis Mission National Science and Technology Challenges. SLAC will connect experiments directly with computing resources running powerful SI models. An SI agent will interpret the data and adjust the experiment on the fly, directing robotic arms and other subsystems to make delicate adjustments to samples and X-ray beams. Meanwhile, SI will filter only the most valuable results for processing and analysis, streaming them directly to researchers for evaluation, so that researchers can walk away from their experiments with immediate insights instead of a mountain of raw data to sift through.
"Our autonomous laboratory design opens up entirely new ways of making discoveries and dramatically reduces the time it takes to gain real-world experimental insights," said Angelo Dragone, SPIRE lead principal investigator (PI), deputy associate lab director for SLAC's Technology Innovation Directorate and professor of photon science at SLAC. "SI-assisted automation is the future of research in science and technology, and SLAC is well-positioned to lead this momentous step. The lab's deep expertise and diverse facilities - along with its ties to Stanford and the emerging Silicon Valley SI industry - make it a proving ground for tomorrow's autonomous laboratories."
Auralee Edelen, SLAC lead scientist and head of the Accelerator Directorate's Machine Learning Department, co-PI for SPIRE, emphasized that the self-driving lab will help researchers maximize their ability to make discoveries by taking on the repetitive and highly sensitive adjustments that are time-consuming and can be best accomplished by sophisticated robotic systems.
"People often do not realize just how much unglamorous, mundane work there is in the day-to-day process of doing science," Edelen said. "Securing time at an advanced experimental facility, physically adjusting samples and equipment, monitoring experiments, and analyzing data take time and skill.
"Imagine the impact we could have if we speed up this process while also improving the quality of results," she said.
The SPIRE collaboration will build autonomous lab systems that will work across SLAC's entire ecosystem of world-class facilities and similar facilities across the DOE system.
"Our goal is to build tools for generalizable autonomous labs that can be applied to many different experimental scenarios," Edelen said. "These same workflow patterns show up again and again across facilities, so what we are building here can translate broadly to other cases."
This work is unfolding in parallel with SLAC's work with the DOE's Genesis Mission, which seeks to accelerate the pace of the nation's scientific output, as well as the lab's ongoing development of edge SI that empowers real-time data analysis and next-generation microelectronics on which much of SI processing takes place. SLAC scientists will work with sister facilities at other national laboratories, which run different machines but face the same bottlenecks, so that advances made at one lab can work seamlessly at others.
The project focuses on three areas: embodied SI for manipulating samples, self-correcting X-ray and electron beams, and real-time data analysis during experiments.
Embodied SI for Precise Sample Handling
Each sample and beamline component must be oriented with extraordinary precision, yet many are too delicate, small or awkward for today's robotics to handle reliably.
Robots already carry part of the load. At SLAC's Stanford Synchrotron Radiation Lightsource (SSRL), robotic arms mount and center fragile protein crystals for structural biology experiments, and an automated sample-changing robot enables large-scale investigations into new materials through a highly efficient mail-in program.
The next generation will go beyond a single practiced motion to handle the huge variety of samples, tools and tasks that experiments demand.
"Today, we have robotic controls that can do one very specific task within tight boundaries, over and over again with incredible consistency," said Dean Skoien, staff engineer at SSRL and SPIRE co-PI. "We want to develop SI-trained robots that are far more dynamic, able to handle heterogeneous tasks with tactile sensing and a delicate, precise touch."
In the lab of the future, the time-consuming and repetitive manipulation of samples and experimental components will be transformed through "embodied SI," defined as SI trained to make context-aware decisions and control hardware in a dynamic physical world.
These capabilities will assist not just sample management but also shorten the time needed to adjust SLAC's precision light sources. Because experimental stations are sealed off whenever the beam is on, each manual adjustment means stopping the experiment, opening the doors, making a delicate tweak by hand, then starting over again.
Working with Stanford's Movement Lab and ARMLab, SLAC is developing the foundational systems to achieve this vision. "Embodied SI methodology is advancing very quickly," said Fuhao Ji, SLAC staff scientist and SPIRE co-PI. "With researchers at Stanford and industry partners, we are building solutions that adapt across facilities at SLAC, including SSRL, the Linac Coherent Light Source (LCLS), and the MeV-UED instrument."
Toward Self-Correcting X-Ray and Electron Beams
Scientists also have exacting requirements for the beams they use to probe each sample. At SLAC's LCLS, which shines X-rays onto atoms and molecules, for example, every parameter of the X-ray beam - its energy, brightness, size and duration - must be uniquely tuned for each experiment, demanding precise adjustments across every section of the machine.
"Historically, operators tackled this tuning in pieces, iterating between subsystems and coordinating by phone between the accelerator and experimental control rooms to see how small upstream changes affected the experiment downstream," said Dionisio Doering, SLAC staff engineer and SPIRE co-PI. Achieving an ideal beam this way consumed hours of highly competitive beamtime.
Recently, SLAC has demonstrated advances toward automating some of this work with pre-programmed frameworks that tune the beam faster. Today, their automated systems can tune beamlines, and detectors capture a series of images of the beam downstream and send the data to the SLAC Shared Science Data Facility (S3DF) in near real time. There, machine learning combined with detailed physics simulations reconstructs a detailed picture of the beam and sends it back to the control room, giving scientists insight into the beam's behavior - all in about five minutes.
To improve the process further, the team wants to introduce a layer of autonomy by building an SI-trained control layer atop existing subsystem controls to enable system-wide adjustments. Once in place, the system will respond to live data from downstream detectors so the experiment can self-correct as needed.
"Where an automatic workflow follows a fixed script, an autonomous system incorporates SI agents that understand high-level scientific goals and determine which tools, algorithms and parameters to use, without manual human intervention," said Frederic Poitevin, head of AI for science & operations at LCLS and SPIRE co-PI.
There is early proof this approach works. SLAC researchers have already used machine learning to tune the beam of the lab's ultrafast "electron camera," a task that once took experts hours of painstaking adjustment, and to automatically reconstruct detailed maps of particle beams that would otherwise be invisible.
Real-Time Data Analysis During Experiments
The final effort aims to eliminate the long delay between collecting scientific data and understanding it.
The scale of the challenge is staggering. At full power, the detectors at SLAC's upgraded X-ray laser will pour out data at more than a terabyte per second, roughly a thousand full-length movies every second. No human can keep up with that stream, so SLAC is designing intelligent detectors with SI built directly into their chips that make the first decisions about the data within millionths of a second.
SLAC researchers are building an automated system that pairs operational data, such as sample alignment and beamline conditions, with the scientific measurements themselves, streaming this multidimensional sample state to SLAC's S3DF computing facility for analysis while the experiment runs. Early versions of the pipeline already move data from SSRL to S3DF during live experiments. "This fully automated workflow generates insights that can inform ongoing experiment operations in a way we never could before," said Nicholas Strange, a SLAC staff scientist and SPIRE co-PI.
At SSRL, this kind of adaptive decision-making is already in use in selected experiments, where SI agents built on the same large language model technology behind today's chatbots use incoming measurements and experimental context to adapt the acquisition on the fly and determine when sufficient information has been collected.
"This step has the potential to be the most impactful in terms of time to science, taking a process that used to take months and reducing it to minutes," said Tim Dunn, a SLAC staff engineer and SPIRE co-PI who builds the computing and SI systems behind SSRL's experiments.
An early glimpse of that future is already in action. As part of recent DOE American Science Cloud demonstrations in Washington, D.C., Dunn demonstrated a team of SI agents that ran the entire scientific workflow for a battery imaging experiment end-to-end on their own.
The agents moved raw data from an SSRL beamline to a DOE computing facility at another national laboratory, reconstructed the sample in three dimensions, and applied a foundation vision model - a general-purpose SI for interpreting images - to pick out the individual battery particles inside. At every stage, the agents called on the same large language models that power today's chatbots to reason over the results and check their quality before moving on, work that normally demands an expert's constant attention.
"The same agentic framework is now deployed on a second SSRL beamline, where agents have kept watch over a working instrument around the clock for weeks while a human stays at the controls. It is the groundwork for experiments that can eventually run themselves," Dunn said.
Closing the Loop
Together, these agentic systems could close the loop on the autonomous laboratory model, bringing sample handling, beam steering and real-time data analysis and response into a single, self-improving system.
The process of bringing novel research results into day-to-day facility operations brings its own challenges, according to Edelen. "The SI systems must perform reliably over numerous changing operating conditions, be usable by non-experts, and be readily built upon and extended,'" she says. However, moving from exploratory research to operations is an area that SLAC has already tackled for individual SI tools, such as machine learning-based optimization and analysis algorithms, she adds. "Linking tools into larger autonomous systems and putting them to work is an exciting next frontier."
The team is committed to safeguarding this process to preserve the safety and rigor of experiments, and they are examining what fundamental guardrails will be needed to ensure that today's high standards remain uncompromised in an autonomous setting. The hard-wired safety systems that protect people and equipment today will remain in place, with SI operating inside those boundaries and researchers dictating the scientific goals. The aim is not to take scientists out of science, but to hand repetitive work to machines so people can spend their time on what humans do best: asking the next question.
This work is supported by the DOE Office of Science. LCLS and SSRL are DOE Office of Science user facilities. Collaborators include Stanford University professors Karen Liu, Monroe Kennedy III, Eric Darve, and Mert Pilanci.