Lawrence Berkeley National Laboratory and University of California, Berkeley
Verified evidenceAutonomous + backstopThat scientists must manually choose every recipe, execute every synthesis step, interpret each XRD result, and plan the next experiment.
Engineering design · Berkeley, California, United States
Collections: Embodied work
Executive brief
Autonomy came from closing the loop between prediction, physical execution, measurement, and replanning; the 37% miss rate shows why scientist validation remains part of the system.
AI value · Target compounds synthesized
Verified36 of 57 target compounds synthesized, a 63% success rate
Research demonstration rather than routine production; 21 targets were not realized and final patterns were manually refined.
Before
Select a target and design a synthesis recipe from literature and thermodynamics. → Dose, mix, heat, characterize, and interpret each experiment.
After
Propose recipes and execute powder dosing, heating, and XRD characterization. → Assess phase yield and propose a new reaction path after failure. → Select target space, audit outputs, manually refine final diffraction patterns, and diagnose system failure modes.
Human boundary
The platform selects follow-up recipes inside the target set; scientists define targets and perform final validation.
Why it matters
That scientists must manually choose every recipe, execute every synthesis step, interpret each XRD result.
Before
How the work ran before the change.
Materials scientist
Select a target and design a synthesis recipe from literature and thermodynamics.
ControlScientific judgment
Laboratory staff
Dose, mix, heat, characterize, and interpret each experiment.
ControlManual lab procedures
What changed
That scientists must manually choose every recipe, execute every synthesis step, interpret each XRD result.
Decision rightAI acts within a human backstop
After
How the same work runs now.
ML recipe system and robots
Propose recipes and execute powder dosing, heating, and XRD characterization.
ControlAir-stable targets and robotic constraints
ML analysis and active learning
Assess phase yield and propose a new reaction path after failure.
ControlTarget-yield threshold
Scientists
Select target space, audit outputs, manually refine final diffraction patterns, and diagnose system failure modes.
ControlHuman scientific accountability
Exception path
Failed targets are inspected for synthesis or computational failure; scientists can adjust decision rules and manually validate diffraction patterns.
Work removed
Decision authority
The platform selects follow-up recipes inside the target set; scientists define targets and perform final validation.
Before
Materials scientist
Select a target and design a synthesis recipe from literature and thermodynamics.
Control: Scientific judgment
Laboratory staff
Dose, mix, heat, characterize, and interpret each experiment.
Control: Manual lab procedures
After
ML recipe system and robots
Propose recipes and execute powder dosing, heating, and XRD characterization.
Control: Air-stable targets and robotic constraints
ML analysis and active learning
Assess phase yield and propose a new reaction path after failure.
Control: Target-yield threshold
Scientists
Select target space, audit outputs, manually refine final diffraction patterns, and diagnose system failure modes.
Control: Human scientific accountability
Work that left the path
Human role before
Scientists and technicians planned and executed each experimental cycle.
Human role after
Scientists define the search space and validate findings while the platform runs repeated synthesis-characterization cycles.
AI roleLiterature-trained recipe generation, XRD interpretation, and active learning connected to robotic execution.
Manual, serial scientist-directed synthesis cycles→36 of 57 target compounds synthesized, a 63% success rate
17 days of continuous closed-loop operation · 353 experiments spanning 33 elements and 40 structural prototypes
Research demonstration rather than routine production; 21 targets were not realized and final patterns were manually refined.
Anti-pattern
Reporting only successful compounds or describing the lab as human-free.
Questions
Portability conditions
Reputation risk
low
Evidence and authority
Current · updated
1 peer reviewed; publication outcomes are verified.
Bundle 1.0.0 · reviewed 2026-08-23 · stable ID 8a2da4a32ef78159
Swiss Plasma Center at EPFL and Google DeepMind · Autonomous + backstop
Operators set plasma objectives while a simulator-trained policy coordinates all coils on physical hardware.
Google · Autonomous + backstop
That AI should only recommend cooling set-point changes for operators to implement manually.
National Grid Electricity Transmission · Autonomous + backstop
Central BVLOS drones/AI automate routine pylon imaging and corrosion grading; engineers decide maintenance.
MIT randomized professional-writing study · Autonomous + backstop
The redesign is a shorter drafting loop with human revision and accountability, not autonomous publication.