Autonomous Quantum Measurement with Agentic AI

Introduction

• QUANTA addresses the need to integrate adaptive control with physical interpretation and troubleshooting across diverse samples and protocols.Transport measurements can probe resistance, Hall and magnetoresistance responses, and quantum oscillations, but reliable insight depends on careful experimental design and interpretation.Source: Introduction · pp. 1–2

Overview

• QUANTA grounds planning in literature, protocols, manuals, and prior data, while a deterministic coordinator validates actions and human researchers retain authorization over physical interventions.The agent interprets returned measurements and faults to revise plans and select subsequent measurements; its actions are bounded and traceable.Source: Introduction · p. 2, Overview · p. 2, Figure 1 · p. 3

Adaptive transport characterization of TaP

• TaP measurements supported metallic temperature dependence, strong field dependence, and nonlinear transverse response, while carrier contributions remained unresolved and the low-temperature upturn depended on zero-field extraction.QUANTA selected field-symmetry and offset analysis before interpreting the intrinsic Hall response, then refined the field sequence to guide follow-up measurements.Source: Adaptive transport characterization of… · p. 4

Autonomous calibration and selection of measurement parameters

• QUANTA recovered 12 readable values after a lock-in overload and selected settings that met resistance-consistency, signal-detectability, and precision criteria.It rejected 3 mA in favor of 1 mA, retained 1 mA when 0.5 mA yielded an SNR of 6.81 below the threshold of 10, and retained longer acquisition after the faster setting missed the precision target.Source: Autonomous calibration and selection… · pp. 3–4, Figure 2 · p. 5

Investigating the nonlinear Hall response in Co2MnGa

• The study tests the nonlinear signal in nominal in-plane I ∥ B geometry because electronic transport and Joule heating can both generate a transverse second-harmonic voltage.The ambiguity matters in Co2MnGa, which exhibits a large anomalous Nernst response from thermal effects.Source: Investigating the nonlinear Hall… · p. 5
• Near-quadratic current scaling establishes a robust second-order response but does not distinguish electronic transport from Joule-heating-induced signals.The fitted exponents are close to two, and the field-odd transverse component also follows approximately I^2.Source: Investigating the nonlinear Hall… · pp. 5–6
• A sign reversal in the imaginary transverse component is not sufficient evidence for distinct mechanisms because a temperature-dependent phase shift can leave the total complex response finite.The response also argues against the simplest transport-scaling picture without establishing a unique microscopic origin.Source: Investigating the nonlinear Hall… · pp. 6–7
• The present dataset cannot reliably separate electronic from thermoelectric contributions, so QUANTA retains both origins as plausible and reports the attribution as inconclusive.The conductance-based fit is strongly ill-conditioned because longitudinal resistance changes by only a few percent over the measured temperatures.Source: Investigating the nonlinear Hall… · p. 7
• Frequency spectra change shape across temperatures after normalization, ruling out simple amplitude rescaling while allowing multiple contributions or one process with evolving dynamics.The spectra are normalized at a common reference frequency.Source: Investigating the nonlinear Hall… · p. 7

Discussion

• QUANTA combines retrieved scientific and laboratory knowledge, specialized agents, adaptive measurement selection, and deterministic safety limits, while physical interventions remain human-authorized.Across TaP and Co2MnGa, it performs adaptive transport characterization and troubleshooting and proposes follow-up measurements for competing mechanisms.Source: Discussion · p. 7
• Current automation does not cover sample mounting, rewiring, or contact changes, and human researchers remain responsible for scientific interpretation and claims.The authors also note that conservative reasoning can make QUANTA reluctant to propose bold hypotheses from incomplete evidence.Source: Discussion · p. 7
• The architecture could extend beyond PPMS transport to other quantum-materials instruments and modalities, with next-measurement selection framed as a way to reduce scientific uncertainty.The discussion names dilution refrigerators, gates, microwave electronics, scanning probes, spectroscopy, and scattering approaches as possible extensions.Source: Discussion · pp. 7–8
• In Co2MnGa, the evidence leaves electronic and thermoelectric origins plausible, with microscopic attribution inconclusive.The conductance-based analysis is ill-conditioned because longitudinal resistance varies by only 3.28% across the measured temperature range.Source: Figure 4 · p. 8

Instrument control and run record

• Measurement requests specify sample, conditions, acquisition, and abort rules; a coordinator validates bounds, command order, and instrument state before dispatch.A persistent worker records raw readings and events, while the researcher handles sample mounting, wiring, and interventions not verifiable through the instrument interface.Source: Instrument control and run record · p. 9
• Run records link plans, data, and interpretations while preserving requested and measured setpoints separately, fault confirmations, retrieval identifiers, analysis versions, and operator overrides.A newly validated request is required before retesting after a fault.Source: Instrument control and run record · p. 9

Quality gates and transport analysis

• Transport acquisition is gated by range, current consistency, complex-signal detectability, and within-window precision before data are accepted for interpretation.The stated thresholds are 3.126 µΩ for current consistency, a complex-signal signal-to-noise ratio of 10, and 5 µΩ for within-window uncertainty.Source: Quality gates and transport analysis · p. 9
• Field symmetry and offset analysis precede intrinsic Hall interpretation; additional procedures address multicarrier-fit uncertainty, quantum oscillations, effective masses, and harmonic responses.The displayed TaP curves remain unsymmetrized, and multicarrier fits are not used to establish carrier count when parameters are non-identifiable.Source: Quality gates and transport analysis · p. 9

Sample fabrication

• The methods require specimen dimensions and contact geometry for resistivity conversion; TaP follows a cited preparation, while Co2MnGa is a 30 nm sputtered film capped with 10 nm MgO and patterned into a Hall bar.The Co2MnGa film was deposited on MgO(001).Source: Sample fabrication · pp. 9–10

Evaluation of autonomous decisions

• The proposed campaign comparison holds starting information, sample access, and stopping criteria constant, and scores qualified completion alongside time, transitions, rejected windows, interventions, and agreement on extracted features.Fault recovery is scored separately from scientific interpretation.Source: Evaluation of autonomous decisions · p. 10

Funding

• Support is acknowledged from DOE Basic Energy Sciences (Award DE-SC0020148) and NSF programs (Awards ITE-2345084 and DMR2522384).Source: Funding · p. 10

1.1 Hardware

• The PPMS DynaCool controls temperature and magnetic field, while external electronics supply current excitation and measure sample voltages for longitudinal or transverse transport.The setup includes a Keithley 6221 current source and two SR860 lock-in amplifiers; voltage-channel assignments depend on sample wiring.Source: §1.1 · pp. 13–14

1.2 Software and data architecture

• QUANTA separates AI scientific coordination from instrument execution: the coordinator proposes plans, while deterministic software validates, executes, and records them within researcher-defined limits.A persistent worker owns live instrument connections; structured plans are checked against a schema and numerical limits before entering a SQLite-backed queue.Source: §1.2 · p. 14
• Durable records retain plans, run state, raw readings, timestamped events, and cached instrument state, including rejected or faulted readings for audit.Plan and configuration hashes link each run to its validated action envelope, and requested set points are stored separately from measured temperature and field.Source: §1.2 · p. 14
• A finite dispatched run may continue after coordinator disconnection, but advancing to a new scientific stage requires the coordinator or a separately validated finite controller.The implementation is persistent at run level, not a generic unattended campaign scheduler.Source: §1.2 · p. 14

2.2 TaP qualification metrics and acceptance criteria

• Acceptance criteria were operational choices, and RMS-current normalization did not independently calibrate the actual sample current.The numerical criteria were not intrinsic material constants or damage thresholds.Source: §2.2 · p. 18
• Early qualification used four blocks of five readings, but blocking did not establish that residual correlation or drift was negligible.The four-block descriptive statistics used the sample standard deviation with three degrees of freedom.Source: §2.2 · pp. 18–19
• The 3-mA TaP field scan was withheld because |∆Y| exceeded GY, but the failed screen did not identify a unique cause.The result was not the difference between the initial and final 1-mA runs.Source: §2.2 · p. 19
• At 0.5 mA, the run passed the half-to-half stationarity screen but failed detectability and percentage-magnetoresistance screens, so it was not qualified as equivalent to the reference.Absence of a resolved mean shift did not establish equivalence to the 1-mA setting.Source: §2.2 · p. 19
• The later +9-T timing study used a prespecified correlation-aware protocol; its conservative correlation heuristic was not a proven bound.Irregular sampling, unresolved positive correlation at lag 10, or failed tail-stationarity checks prevented qualification.Source: §2.2 · p. 20
• All 30 timing windows remained inconclusive under the full acceptance rules; independent follow-up was not completed, and no net time saving was established.A small within-window uncertainty alone did not qualify a faster setting.Source: §2.2 · p. 20

2.3 Structured decision summaries

• Structured decision records separate observations, interpretations and proposals; a successful remeasurement can clear an overload without uniquely diagnosing its cause.The instruction hierarchy also distinguishes campaign records from machine-readable executable values.Source: §2.1 · p. 18, §2.3 · p. 20
• At 20 K, Co2MnGa measurements supported a predominantly field-odd quadratic voltage response, but did not resolve its microscopic attribution.The reported power-law exponent was p = 2.034±0.056 for the 0.5–1 mA amplitude series.Source: §2.3 · pp. 20, 22

3 Experimental Examples of Agent-Guided Scientific Decisions

• QUANTA’s examples demonstrate adaptive planning and controlled investigation of unexpected responses, with physical interventions kept under human control.Scientific planning was separated from validated instrument execution.Source: §3 · p. 22

3.1 Adaptive measurement planning and run closure

• After an 85-point-per-branch 300 K survey, the agent shifted to 25 points with six repeats per point across the full field span.Follow-up measurements were acquired at 100, 20 and 2 K; the researcher selected the 2 K endpoint.Source: §3.1 · p. 22

3.2 Reference-assisted troubleshooting of an unexpected response

• An initial 100–200–100 nA comparison showed an approximately 4% response change, prompting a reference-resistor test.With the narrower 2-µA range, the observed variation fell within the displayed ±1% margin; evidence supported retaining 100 nA, while 200 nA was not qualified.Source: §3.2 · p. 22
• Reference measurements showed configuration dependence but did not establish the initial shift’s microscopic or instrumental origin.Sample and resistor exchanges required human intervention.Source: §3.2 · p. 22

4.1 Layered safeguards

• Before dispatch, software validates plan fields and verifies configuration and worker state; during execution, it monitors instrument status and attempts source-off on completion, pause, abort or fault.The resulting source readback is recorded, but wiring, contact continuity, sample mounting and true sample current are not inferred from telemetry.Source: §4.1 · p. 22
• Co2MnGa measurements used a revised protocol of 25 points and six repeats per point at 100, 20 and 2 K, following an initial 85-point-per-branch sweep at 300 K.The run recorded source-off and environmental return checks; puck exchange remained pending.Source: Figure 4 · p. 23
• A reference-resistor comparison found a similar response on the 20-µA range and substantially reduced variation on the 2-µA range; the sequence supported retaining 100 nA, not qualifying 200 nA or uniquely identifying the initial response's origin.A human operator performed the physical connections.Source: Figure 5 · p. 24
• Hard limits, authorized qualification envelopes and demonstrated ranges are distinct; a hardware rating or validator ceiling alone does not establish that a setting is demonstrated.Short, newly validated pilots may be attempted only within the authorized qualification envelope.Source: §4.1 · p. 25
• Campaign-specific numerical configurations are auditable records, not transferable settings or evidence that all extremes were used together or represent intrinsic damage thresholds.Source: §4.1 · p. 25

4.2 Fault classification and response

• Only changes within the recorded envelope can be tested automatically; branches requiring rewiring, sample exchange or contact repair are held for a person.The workflow distinguishes electrical protection, acquisition validity and scientific precision.Source: §4.2 · p. 25

• The dashboard is a read-only projection of committed campaign records, not an instrument-control path.Source: §5 · p. 25
• It displays campaign progress, cached instrument state, pending validated plans, data-quality annotations and versioned analysis products.Source: §5 · p. 25
• Run branches and repeats remain separable, while unverified telemetry is marked historical or unverified and estimated completion time is withheld.Display metadata cannot select an instrument endpoint or authorize dispatch.Source: §5 · p. 25

• The archive contains 333 native sequence files, with multi-label command counts spanning temperature, field, transport, thermal transport, rotation and heat-capacity procedures.Counts include 277 temperature-control, 212 field-control and 165 native resistance or transport files; a file can contribute to multiple modality counts.Source: §6 · p. 25
• Path matching found 52 TaP and 8 Co2MnGa sequences, so the campaigns have same-material precedents and are not material-held-out tests of generalization.The corpus can provide procedural evidence, but missing material-name matches do not rule out indirect protocol leakage.Source: §6 · p. 25
• A prospective benchmark should split by specimen, material family, contributor and time, remove near-duplicate protocols, log retrieval identifiers and hashes, and report results with and without same-material precedents.No prespecified held-out evaluation was completed; the sequence analysis documents coverage and leakage risk, not a generalization score.Source: §6 · p. 25

7 Reproducibility, Provenance and Scope

• Both campaigns used OpenAI Astra Ultra for the Codex coordinator and specialist agents supporting planning, troubleshooting and data interpretation.Source: §7 · p. 28

7.1 Evidence archive and interpretive limits

• The evidence archive preserves control software, configurations, tests, sequences, campaign records, run databases, exports, analyses, dashboard snapshots and reports.Figure-generating analyses are kept outside the source-data tree and read the archive without modifying campaign records.Source: §7.1 · p. 28
• Sequential readbacks, programmed rather than independently measured current, and the lack of dimensions and contact calibration constrain what the archived measurements establish.PPMS stability alone does not establish sample equilibrium; the corpus audit also does not replace a prospective held-out benchmark.Source: §7.1 · p. 28

7.2 Agent activity and computational usage

• Co2MnGa accounted for 72.4% and TaP 18.2% of experimental-campaign token usage.Campaign monitoring, specialist review and implementation, measurement planning and execution, and instrument diagnosis and qualification were among the largest expenditure categories.Source: §7.2 · p. 28
• Recorded usage totaled approximately 2.75 billion tokens: 97.5% cached input, 2.3% uncached input and 0.2% output.These cumulative counts include development and support activities and are not measures of independent reasoning, unique input content or experimental efficiency.Source: §7.2 · p. 28