Authors: Eunbi Rha, Abhijatmedhi Chotrattanapituk, Mouyang Cheng, Chung-tao Chou, Shenyong Ye, Ashley S. Dale, Long Ju, Luqiao Liu, Mingda Li
Organizations: Quantum Measurement Group, MIT, Cambridge, MA 02139, USA · Department of Nuclear Science and Engineering, MIT, Cambridge, MA 02139, USA · Department of Electrical Engineering and Computer Science, MIT, Cambridge, MA 02139, USA · Center for Computational Science & Engineering, MIT, Cambridge, MA 02139, USA · Department of Materials Science and Engineering, MIT, Cambridge, MA 02139, USA · Department of Physics, MIT, Cambridge, MA 02139, USA
Quantum materials exhibit diverse electronic and functional phenomena, but their characterization requires extensive experimental expertise, iterative decision-making and accumulated laboratory knowledge that is difficult to document and transfer. Here we introduce Quantum Agent (QUANTA), an agentic artificial intelligence (AI) framework for autonomous quantum-materials characterization. Using a widely adopted Physical Property Measurement System (PPMS) as an experimental platform, QUANTA combines retrieval augmentation using scientific literature, experimental protocols and prior measurements with specialized agents for experimental planning, troubleshooting, and data interpretation. A separate deterministic control layer validates instrument actions. Guided by incoming data, QUANTA evaluates measurement quality, revises experimental plans and selects subsequent measurements. We demonstrate these capabilities through transport measurements of the topological Weyl semimetal TaP, as well as investigations of nonlinear responses in magnetic material Co2MnGa. These results establish QUANTA as a generalizable framework for evidence-driven, safeguarded and closed-loop quantum-materials characterization.
Figures & tables
Figure 1: Architecture of QUANTA for autonomous measurement of quantum materials. A human researcher (top yellow block) supplies the scientific question and authorizes physical interventions. The main agent (purple block) iterates through evidence review, measurement planning and interpretation, drawing on literature, protocols, manuals and prior data in a shared knowledge store; specialist agents support these tasks. Plans pass to an execution coordinator (central blue block) for validation before a dedicated worker acts through the PPMS and measurement computers. Inside the DynaCool (bottom left), the PPMS sets temperature and field, while external current source and lock-in instruments provide excitation and electrical readout. Instrument data and faults return to the agent for the next decision. The dashed yellow arrow denote the researcher’s physical oversight of the entire measurement process.
Figure 2: Autonomous calibration of transport-measurement parameters by QUANTA. a. Instrument-range recovery. A 100- μ A root mean square (RMS) excitation at 300 K and 0 T produces a lock-in overload. QUANTA increases the lock-in input headroom, after which 12 readable measurements are recovered. The panel also shows the detailed closed-loop workflow, including fault reporting, agent reasoning, proposal of revised settings, coordinator validation and execution of the corrected measurement. b. Selection of excitation current for a consistent resistance measurement. At 100 K and 0 T, the resistance inferred at 3 mA differs from the settled 1-mA reference. QUANTA therefore rejects the 3-mA condition and retains 1 mA, for which all readings pass the validity checks. c. Selection of excitation current for high signal-to-noise ratio (SNR). At 2 K and 0 T, reducing the current to 0.5 mA gives a low complex-signal SNR of 6.81, below the acceptance threshold of 10. QUANTA therefore retains the higher-current condition. d. Selection of acquisition time duration. At 2 K and +9 T, 1-s and 3-s lock-in time constants are evaluated over increasing averaging spans against a 5- μΩ within-window uncertainty target. The faster acquisition does not satisfy the required precision, so QUANTA retains the longer measurement condition.
Figure 3: Adaptive transport characterization of TaP. The user question (yellow) and retrieved knowledge (grey) define candidate transport hypotheses and motivate an initial measurement survey ( steps 1–2 ). Temperature- and field-dependent resistance measurements are acquired through proposed sequences, and the resulting trends are used to select appropriate analyses ( step 3 ) and refine the next measurement sequence ( step 4 ). The example data show zero-field resistance versus temperature together with raw transverse and longitudinal field-dependent resistance in TaP, without transverse anti-symmetrization or longitudinal symmetrization. The transverse curves retain a substantial field-even contribution. Checked boxes denote agent-selected actions. Refined measurements are returned for execution, after which the agent summarizes supported observations and remaining uncertainties ( step 5 ) and formulates a final response that separates candidate interpretations from established findings ( step 6 ).
Figure 4: Testing competing origins of the nonlinear transverse voltage in Co 2 MnGa. QUANTA progressively tests the mechanism assignment through four questions. Q1 , current sweeps from 0.1–1 mA establish an approximately quadratic response, with fitted exponents close to 2, but this scaling is compatible with both electronic and Joule-heating-induced signals. Q2 , temperature-dependent measurements reveal a sign reversal in the imaginary transverse component, although the full complex response remains finite and the reversal alone does not establish distinct mechanisms. Q3 , frequency sweeps over 17.777–3555.4 Hz show temperature-dependent spectral shapes after normalization at 177.77 Hz, ruling out simple rescaling of a fixed response while remaining compatible with either multiple contributions or a single process with evolving dynamics. Q4 , conductance-based scaling is strongly ill-conditioned because the longitudinal resistance spans only 3.28% across the measured temperature range, preventing reliable separation of candidate contributions. QUANTA therefore retains both electronic and thermoelectric origins as plausible and reports the microscopic attribution as inconclusive.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Supplementary Fig. 1: Physical configuration of the agentic PPMS experimental platform. (a) Laboratory overview showing the PPMS control computer, DynaCool, external measurement rack, module switch, and measurement computer. (b) Close-up of the external measurement electronics, including the Keithley 6221 current source, two SR860 lock-in amplifiers, and sample wiring box. (c) DynaCool sample-interface region and adjacent connection hardware. The photographs identify the physical components corresponding to the control and measurement paths in main-text Fig. 1; individual sample-contact assignments depend on the experimental wiring.
Supplementary Fig. 2: Scientific workflow and instrument connectivity of QUANTA. (a) Agent-guided measurement workflow, distinguishing researcher-defined objectives and physical verification, AI-based scientific coordination, stored knowledge and campaign records, and deterministic execution software. The AI coordinator retrieves evidence, formulates measurement plans, and interprets results, while software validation and a persistent worker enforce the authorized operating envelope and record experimental outcomes. (b) Connectivity among the researcher’s computer, measurement computer, PPMS control computer, and physical instruments. The AI coordinator communicates with the MCP service over SSH, while the persistent worker interfaces with the PPMS through MultiPyVu and with the external electrical instruments through NI-VISA/GPIB. SQLite storage supports persistent execution records, numerical analysis, and a read-only dashboard. Arrows distinguish software communication, stored records, electrical signals, and the sample environment.
Supplementary Fig. 3: Detailed software architecture, execution safeguards and data provenance. The implementation-level architecture underlying the workflow in Supplementary Fig. 2 . (a) The researcher defines the objective, operating envelope and stop authority, while the AI coordinator retrieves evidence and selects measurements. (b) Structured requests pass through application-level validation and a durable queue before execution by the persistent worker, which maintains exclusive instrument ownership. Deployment checks, stop controls, instrument-state safeguards and immutable run records support bounded execution and auditability. (c) NI-VISA/GPIB carries instrument commands and readbacks, separately from excitation, voltage and phase-reference signal paths. (d) MultiPyVu connects the worker to MultiVu and the DynaCool. Run-level persistence does not constitute an unattended campaign scheduler; source-off attempts do not zero the magnet or cancel retained PPMS targets.
Layer
Required content and role
Researcher intake
Objective, sample identity, physical wiring evidence, allowed temperature, field and excitation range, time budget, completion criterion, excluded physical actions and stop authority. Establishes campaign scope but does not encode an executable command
Repository entry instructions
Exact deployment, worker owner, configuration, state directory and requirement to read the operating guide. Prevents reuse of a stale endpoint or a different sample’s profile
Operating guide
Manual and literature retrieval, sequence search, qualification, validation, monitoring, troubleshooting and shutdown policy. Defines how a scientific objective is converted into bounded stages
Campaign record and checklist
Hypotheses, controls, authorization, limits, accepted and rejected evidence, pending route and final-state requirement. Maintains the evolving scientific state and handoff record
Plan and deployment configuration
Ordered set points, rates, electrical settings, repetitions, stop conditions, numerical ceilings and configuration hash. Supplies the only representation accepted by the validator and worker
SQLite run record
Plans, runs, readings, events, cached state, timestamps, quality flags and shutdown result. Connects a scientific decision to the exact action and returned evidence
Appendix
Supplementary Table 1: Instruction and record hierarchy used by QUANTA.
Case
Evidence, guarded action and interpretation limit
TaP overload, 300 K, 0 T
Evidence: At 100μA RMS and 177.77Hz , the SR860 overloaded with a 10mV input range and 10μV sensitivity; no reading was accepted. Action: Stop excitation, confirm source-off and retest in a new validated plan with a 1V range and 50mV sensitivity. Conclusion: The retest produced 12/12 readable values, establishing recovery but not the physical overload cause
TaP current bracket, 100 K, 0 T
Evidence: Within a 1–3–1 mA RMS bracket, the 3-mA run and subsequent 1-mA return differed by 4.157μΩ in Y/I . Their combined descriptive block uncertainty was 1.042μΩ , giving a three-times-uncertainty threshold of 3.126μΩ , which exceeded the alternative 1% signal-scale floor. The observed difference exceeded this threshold. Action: Withhold the 3-mA field scan and return to a longer-settled 1-mA plan. Conclusion: The 1-mA scan yielded 180/180 valid readings; heating, non-ohmic response, contact change, pickup and drift were not separated, and 3-mA damage was not established
TaP low-current bracket, 2 K, 0 T
Evidence: In the 1–0.5–1 mA RMS comparison, the 0.5-mA run gave SNRcomplex=rˉX2+rˉY2/uX2+uY2=6.81 , where rˉX and rˉY are the mean current-normalized components and uX and uY are their four-block standard uncertainties. This fell below the selected detectability cutoff of 10. Action: Do not promote 0.5 mA as an equivalent high-field setting. Conclusion: No mean shift was resolved, but equivalence remained unproven because detectability failed
TaP timing test, 2 K, +9 T
Evidence: Four 61-reading traces compared 1-s and 3-s time constants. Two traces per setting, three candidate delays for 1 s, two for 3 s and three window sizes produced 2×3×3+2×2×3=30 analysis windows. The selected precision target was utotal,X=uwithin,X2+sbetween,X2≤5μΩ , subject to validity, correlation, batch-count, stationarity and repeatability checks. This target was a standard uncertainty, not a 95% interval half-width. Action: Retain all traces and withhold a faster production setting until all acceptance criteria pass. Conclusion: All 30 windows were inconclusive; unresolved repeatability was not treated as zero, and neither a timing setting nor a net time saving was qualified
Appendix
Supplementary Table 2: Examples of structured evidence-to-action decisions. TaP qualification metrics and acceptance criteria are defined in Section 2.2 .
Supplementary Fig. 4: Adaptive measurement planning and run closure in Co 2 MnGa. (a) Experimental decision sequence, from review of the initial 300 K survey to revised sampling, temperature-dependent acquisition and final-state checks. Purple denotes AI review and planning, blue denotes software execution, and yellow identifies human decisions or interventions. (b) Initial 300 K out-of-plane field sweep, acquired with 85 points per branch. (c–e) Subsequent measurements at 100, 20 and 2 K using the revised sampling protocol of 25 points and six repeats per point. Filled circles and open squares distinguish opposite field-sweep directions. (f) Zero-field resistance during the temperature sequence and final return to 300 K. The run concluded with recorded source-off and environmental return checks; the physical puck exchange remained pending.
Supplementary Fig. 5: Reference-assisted investigation of an unexpected transport response in Co 2 MnGa. (a) Diagnostic workflow initiated by an approximately 4% change in the measured in-phase response. The agent proposed a reference-resistor comparison and evaluated the results, while a human operator performed the required physical connections. Purple indicates AI review and planning, blue indicates software acquisition, and yellow indicates human intervention. (b) Initial sample measurement using the 20- μ A current-source range. (c) Reference-resistor measurement with the same range, exhibiting a comparable response. (d) Reference-resistor measurement using the 2- μ A range, showing a substantially reduced variation. (e) Sample remeasurement after restoring the original wiring and using the 2- μ A range. Blue circles and orange squares represent the two lock-in channels during 100–200–100 nA RMS excitation comparisons. The shaded region denotes the ±1% comparison margin. The diagnostic sequence supported retaining 100 nA but did not qualify 200 nA or uniquely establish the origin of the initial response.
Campaign configuration
Recorded action envelope and interpretation
TaP AC transport
Environment: 2–301 K and ∣B∣≤9\text{,}\mathrm{T} ; maximum rates $2\text{\,}\mathrm{K}\text{\,}{\mathrm{min}}^{-1}$ and $0.01\text{\,}\mathrm{T}\text{\,}{\mathrm{s}}^{-1}$ . Electrical: $I_{\mathrm{peak}}\leq$4.3\text{\,}\mathrm{mA} , compliance ≤0.1\text{,}\mathrm{V}$$ and configured estimated power ≤10μW . Run bound: 3,000 points, 50,000 readings and 24 h. The current ceiling was a conservative qualification boundary, not a demonstrated sample-safe limit
Co 2 MnGa out-of-plane Hall
Environment: 2–300 K and ∣B∣≤9\text{,}\mathrm{T} ; maximum rates $3\text{\,}\mathrm{K}\text{\,}{\mathrm{min}}^{-1}$ and $0.02\text{\,}\mathrm{T}\text{\,}{\mathrm{s}}^{-1}$ . Electrical: $I_{\mathrm{peak}}\leq 0.30\,\mu\mathrm{A}$ , compliance $\leq$0.1\text{\,}\mathrm{V} and configured estimated power ≤10\text{,}\mathrm{nW}$$ . Run bound: 10,000 points, 100,000 readings and 24 h. Both voltage channels were read sequentially; a transverse channel label did not by itself establish Hall polarity
Co 2 MnGa in-plane nonlinear campaign
Environment: 2–300 K and ∣B∣≤1\text{,}\mathrm{T} ; maximum rates $4\text{\,}\mathrm{K}\text{\,}{\mathrm{min}}^{-1}$ and $0.01\text{\,}\mathrm{T}\text{\,}{\mathrm{s}}^{-1}$ . Electrical: $I_{\mathrm{peak}}\leq$1.5\text{\,}\mathrm{mA} , compliance ≤5\text{,}\mathrm{V} and port-power screening $\leq$15\text{\,}\mathrm{mW} ; commissioned transport used at most 1mA RMS with a 180s source-on timer. Run bound: 10,000 points, 100,000 readings and 6 h. Port-power screening was not a measurement of sample or contact dissipation
Appendix
Supplementary Table 3: Selected deployed action envelopes and experiment-specific gates.
Observation class
Response and remaining limitation
Stale deployment, owner mismatch or unverified wiring map
Hold dispatch; reconcile worker, configuration and physical record. Unresolved: actual lead continuity, grounding and sample identity require independent evidence
Malformed or out-of-envelope plan
Reject before queueing and return the failed field or numerical bound. Scientific desirability cannot override a hard limit
Source compliance, overtemperature or inconsistent source state
Stop the affected run, preserve flags and attempt verified source-off. Unresolved: programmed current is not an independent measurement of sample current
SR860 input overload or reference loss during excitation
Reject affected readings, stop when required and use a new validated diagnostic plan. Unresolved: a range change can restore acquisition without identifying the physical cause
Chamber fault, non-finite or persistently unknown T/B state
Pause or fault, attempt source-off and retain the last PPMS targets for explicit review. Unresolved: chamber mode is not a pressure measurement; electrical shutdown does not zero the magnet
Communication or worker fault
Preserve durable run state, block blind restart and reconcile ownership before a new run. Unresolved: process or bus failure can prevent a guaranteed shutdown; this is not an emergency-stop system
Appendix
Supplementary Table 4: Fault and data-quality classes used in the supervisory workflow.
Supplementary Fig. 6: Read-only campaign dashboard. Example Co 2 MnGa transport workspace showing the campaign objective, recorded-run progress, roadmap, cached temperature/field/source state, measurement tabs, individual field branches and provenance controls. The visible “telemetry unverified” and “estimate withheld” messages illustrate the freshness policy: the interface does not present historical cache values as a verified live instrument state. The dashboard has no instrument-control path.
Supplementary Fig. 7: Composition of the laboratory sequence corpus. (a) Number of .seq files by contributor directory. (b) Multi-label modality counts based on active command tokens; one file may contribute to several bars. (c) Direct case-insensitive material-name matches in sequence paths for the four manuscript examples. TaP and Co 2 MnGa therefore require explicit same-material exclusion in a held-out evaluation.
Supplementary Fig. 8: Agent activity and token usage in the PPMS project. (a) Cumulative token consumption by task category, including campaign monitoring and coordination, specialist review and implementation, measurement planning and execution, instrument diagnosis and qualification, and supporting activities. Values are reported in millions of tokens. (b) Allocation of total token usage across the Co 2 MnGa (CMG) and TaP, together with shared system and review activities. (c) Aggregate token composition, comprising 2.753 billion cached-input tokens (97.5%), 64.78 million uncached-input tokens (2.3%) and 7.00 million output tokens (0.2%). Colors distinguish cached input, uncached input and output across all panels. The statistics represent cumulative recorded model usage, including development and supporting activities, rather than independent reasoning steps or direct measures of scientific performance.
While modern Large Language Models (LLMs) and agentic artificial intelligence (AI) have demonstrated transformative capabilities in digital domains, the realization of embodied AI capable of real-world scientific discovery remains a difficult frontier. The advancements are hindered by the inherent complexity of integrating high-level reasoning, multimodal information processing and real-time physical execution. Here we introduce Qumus, the first AI quantum materials experimentalist. Physically embodied within a robotic mini-laboratory, Qumus is an intelligent, multimodal, and multi-agent system designed for the creation and nano-processing of atomically thin two-dimensional (2D) materials and stacked van der Waals (vdW) structures. Qumus autonomously navigates the full scientific cycle, from hypothesis generation and protocol planning to multi-step experimental execution, result analysis and reporting, acting as an experimentalist. Markedly, the system has achieved, for the first time, the AI-creation of graphene, as well as the first AI-fabrication of complex nanodevices including atomically thin field-effect transistors via vdW stacking. Qumus excels at these tasks by demonstrating autonomous error correction and closed-loop experimentation. Our results establish a generalizable framework for self-improving embodied AI systems that learn directly from the quantum world, opening a pathway toward accelerated discovery in quantum materials, electronics and beyond.
Lihan Shi, Zhaoyi Joy Zheng, Xinzhe Juan +14
Department of Physics, Princeton University, Princeton, New Jersey 08544, USA · Department of Electrical and Computer Engineering, Princeton University, Princeton, New Jersey 08544, USA · Department of Computer Science and Engineering, University of Michigan, Ann Arbor, Michigan 48109, USA +5
Large language model agents have begun to carry out laboratory measurements, but their scientific conclusions are useful only if they hold up under scrutiny. Demonstrating this requires controlled studies because models are stochastic and the context a model is given steers its conclusions. Here we introduce HERA, an autonomous agent designed to carry out cryogenic transport measurements of quantum devices. HERA draws on a knowledge base curated by lab members that holds measurement protocols, software tools, and literature references. We assessed HERA with a blinded trial against an expert on the stacking order of two multilayer graphene devices, an ablation of its knowledge base, and hundreds of offline replays in which models designed the trial's central measurement without running it. HERA's conclusions matched the expert's sealed conclusions on both tested devices. The ablation and the replay studies measured the impact of the knowledge base across different models, showing that with it every model designed better measurements and reasoned more soundly. To test how readily HERA extends to a new problem, we tasked it with finding and characterizing superconductivity in a different material system. This required a few hours of expert time to develop a new experimental protocol, after which HERA did both autonomously. These results show that, with a knowledge base designed for the task, language model agents are capable of sound experimental science.
Zachary W. Gomez, Nathan Bittner, Jordan T. McCourt +7
Stanford Institute for Materials and Energy Sciences, SLAC National Accelerator Laboratory, Menlo Park, CA 94025 · Department of Physics, Stanford University, Stanford, CA 94305 · Research Center for Electronic and Optical Materials, National Institute for Materials Science, 1-1 Namiki, Tsukuba 305-0044, Japan +1
We implement an agentic AI workflow built around a large language model (LLM) agent for autonomous experiments with nitrogen-vacancy (NV) centers in diamond. NV centers are a widely used platform for quantum sensing, and the ability to control many measurements from a computer makes NV experiments a natural setting for autonomous workflows. We make two main contributions. First, we demonstrate an autonomous NV experiment workflow that combines persistent project records, quantitative calculation and data analysis tools, and deterministic experiment control. In one autonomous experiment, the agent selected a single NV center, calibrated its resonant frequency, measured T2∗ with Ramsey measurements, and added a Carr--Purcell--Meiboom--Gill (CPMG) measurement to check a weak feature that could be related to nearby 13C. Second, we introduce two offline benchmarks that evaluate the agent's reasoning separately from laboratory execution. We evaluated both benchmarks with GPT-5.4, GPT-5.5, and GPT-5.6 Sol. In the Ramsey checkpoint benchmark, greater reasoning effort generally improved recognition of a residual resonance calibration offset. By contrast, in the pulsed optically detected magnetic resonance (pODMR) data evaluation benchmark, pulse sequence information alone produced more false positive resonance judgments at higher reasoning effort. Requiring an expected signal calculation kept false positive rates low across all three models and reasoning settings. The results suggest a clear division of labor for autonomous experiments. The agent forms scientific hypotheses and uses quantitative tools to evaluate data, while deterministic code controls the hardware and enforces safety constraints.
Department of Nuclear Science and Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, USA · Department of Chemistry, Massachusetts Institute of Technology, Cambridge, MA 02139, USA · Department of Physics, Massachusetts Institute of Technology, Cambridge, MA 02139, USA